Systems and methods for real-time multi-document capture using machine learning models

A machine learning system for real-time multi-document capture addresses inefficiencies in existing scanning by dynamically detecting and segmenting multiple documents, preventing duplicates, and enhancing image quality, resulting in a seamless and efficient scanning experience.

WO2026241090A1PCT designated stage Publication Date: 2026-11-19GOOGLE LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027332
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-13
Filing Date
2026-05-08
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing mobile document scanning solutions are limited to single-document capture, struggle with real-time multi-document detection, and lack effective duplicate prevention and image quality enhancement, leading to inefficient and error-prone workflows.

Method used

A machine learning-based system for real-time multi-document capture that dynamically detects and segments multiple documents within a single frame, prevents duplicates through hierarchical feature comparison, and enhances image quality during capture.

Benefits of technology

Enables seamless, intuitive multi-page scanning with reduced duplicates and improved image quality, streamlining the capture process and reducing computational overhead.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027332_19112026_PF_FP_ABST
    Figure US2026027332_19112026_PF_FP_ABST
Patent Text Reader

Abstract

An example method includes identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents. The method also includes detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document. The method further includes selecting, by the machine learning model and based on the respective set of comers, one or more documents, of the plurality of candidate documents, that are to be digitally scanned. The method also includes scanning, based on the respective set of corners, the one or more selected documents. The method also includes providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.
Need to check novelty before this filing date? Find Prior Art

Description

Atty. Docket: 25-0667-WOSYSTEMS AND METHODS FOR REAL-TIME MULTI-DOCUMENT CAPTURE USING MACHINE LEARNING MODELSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This present application claims priority to Indian provisional application no.202511045841 filed May 13, 2025, the full disclosure of which is incorporated herein by reference in its entirety.BACKGROUND

[0002] Image capturing devices can be used to capture images of documents. Users generally review the captured document prior to saving it, recapturing the document, and / or capturing another document. Users may also apply manual adjustments to captured documents to optimize image quality. Manual adjustments can include determining document corners and / or boundaries, improving image resolution, determining duplicate documents, and so forth.SUMMARY

[0003] This application is generally directed to mobile document scanning approaches that involve a fast, intuitive, and intelligent mobile scanning system capable of handling multi-page scenarios seamlessly through real-time, simultaneous multi-document detection, real-time duplicate prevention, and real-time image quality enhancement within a continuous image capture experience. Also described herein is a motion metering system to dynamically control frame capture timing and determine whether to capture a current frame.

[0004] In one aspect, a computer-implemented method is provided. The method includes identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents. The method also includes detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document. The method further includes selecting, by the machine learning model and based on the respective set of corners, one or more documents, of the plurality of candidate documents, that are to be digitally scanned. The method also includes scanning, based on the respective set of corners, the one or more selected documents. The method also includes providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

[0005] In a second aspect, a device is provided. The device includes one or more processors and data storage. The data storage has stored thereon computer-executable instructions that,Atty. Docket: 25-0667-WOwhen executed by one or more processors, cause the device to carry out functions. The functions include: identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents; detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document; selecting, by the machine learning model and based on the respective set of comers, one or more documents, of the plurality of candidate documents, that are to be digitally scanned; scanning, based on the respective set of corners, the one or more selected documents; and providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

[0006] In a third aspect, a computer program is provided. The computer program includes instructions that, when executed by a computer, cause the computer to carry out functions. The functions include: identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents; detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document; selecting, by the machine learning model and based on the respective set of comers, one or more documents, of the plurality of candidate documents, that are to be digitally scanned; scanning, based on the respective set of corners, the one or more selected documents; and providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

[0007] In a fourth aspect, an article of manufacture is provided. The article of manufacture includes one or more computer readable media having computer-readable instructions stored thereon that, when executed by one or more processors of a device, cause the device to carry out functions. The functions include: identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents; detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document; selecting, by the machine learning model and based on the respective set of corners, one or more documents, of the plurality of candidate documents, that are to be digitally scanned; scanning, based on the respective set of corners, the one or more selected documents; and providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

[0008] In a fifth aspect, a system is provided. The system includes: means for identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents; means for detecting, by a machine learning model and forAtty. Docket: 25-0667-WOeach of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document; means for selecting, by the machine learning model and based on the respective set of corners, one or more documents, of the plurality of candidate documents, that are to be digitally scanned; means for scanning, based on the respective set of corners, the one or more selected documents; and means for providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

[0009] In a sixth aspect, a computer-implemented method is provided. The method includes detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest. The method also includes tracking the one or more features in consecutive frames of the one or more image frames. The method additionally includes determining magnitudes of each motion vector corresponding to the tracked one or more features. The method also includes determining an aggregate motion metric for the magnitudes of each motion vector. The method additionally includes determining a frame stability parameter based on the aggregate motion metric. The method further includes triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

[0010] In a seventh aspect, a device is provided. The device includes one or more processors and data storage. The data storage has stored thereon computer-executable instructions that, when executed by one or more processors, cause the device to carry out functions. The functions include: detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest; tracking the one or more features in consecutive frames of the one or more image frames; determining magnitudes of each motion vector corresponding to the tracked one or more features; determining an aggregate motion metric for the magnitudes of each motion vector; determining a frame stability parameter based on the aggregate motion metric; and triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

[0011] In an eighth aspect, a computer program is provided. The computer program includes instructions that, when executed by a computer, cause the computer to carry out functions. The functions include: detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest; tracking the one or more features in consecutive frames of the one or more image frames; determining magnitudes of each motion vector corresponding to the tracked one or more features; determining an aggregate motion metric for the magnitudes of each motion vector; determiningAtty. Docket: 25-0667-WOa frame stability parameter based on the aggregate motion metric; and triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

[0012] In a ninth aspect, an article of manufacture is provided. The article of manufacture includes one or more computer readable media having computer-readable instructions stored thereon that, when executed by one or more processors of a device, cause the device to carry out functions. The functions include: detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest; tracking the one or more features in consecutive frames of the one or more image frames; determining magnitudes of each motion vector corresponding to the tracked one or more features; determining an aggregate motion metric for the magnitudes of each motion vector; determining a frame stability parameter based on the aggregate motion metric; and triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

[0013] In a tenth aspect, a system is provided. The system includes: means for detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest; means for tracking the one or more features in consecutive frames of the one or more image frames; means for determining magnitudes of each motion vector corresponding to the tracked one or more features; means for determining an aggregate motion metric for the magnitudes of each motion vector; means for determining a frame stability parameter based on the aggregate motion metric; and means for triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

[0014] The foregoing summary is illustrative only and is not intended to be in any way limiting. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features will become apparent by reference to the figures and the following detailed description and the accompanying drawings.BRIEF DESCRIPTION OF THE FIGURES

[0015] FIG. 1 is an example overview of a multi-document detector system, in accordance with example embodiments.

[0016] FIG. 2 illustrates examples of multi-document detection, in accordance with example embodiments.

[0017] FIG. 3 illustrates an example machine learning model for multi -document scanning, in accordance with example embodiments.Atty. Docket: 25-0667-WO

[0018] FIG. 4 illustrates an example of corner detection, in accordance with example embodiments.

[0019] FIG. 5 illustrates another example machine learning model for multi-document scanning, in accordance with example embodiments.

[0020] FIG. 6 illustrates an example multi-document scanning system, in accordance with example embodiments.

[0021] FIG. 7A illustrates an example multi-document scanning process, in accordance with example embodiments.

[0022] FIG. 7B illustrates an additional multi -document scanning process of FIG. 7A, in accordance with example embodiments.

[0023] FIG. 8A illustrates another example of a multi -document scanning process, in accordance with example embodiments.

[0024] FIG. 8B further illustrates the example multi -document scanning process of FIG. 8A, in accordance with example embodiments.

[0025] FIG. 9 illustrates an example multi-document scanning process, in accordance with example embodiments.

[0026] FIG. 10 illustrates an example unwarping of a scanned document, in accordance with example embodiments.

[0027] FIG. 11 illustrates example post-processing of a scanned document, in accordance with example embodiments.

[0028] FIG. 12 illustrates an example shadow removal of a scanned document, in accordance with example embodiments.

[0029] FIG. 13 illustrates an example document deduplication process, in accordance with example embodiments.

[0030] FIG. 14 further illustrates the document deduplication process of FIG. 13, in accordance with example embodiments.

[0031] FIG. 15 illustrates an example user interface for document deduplication, in accordance with example embodiments.

[0032] FIG. 16 illustrates generation of an example synthetic training dataset, in accordance with example embodiments.

[0033] FIG. 17 illustrates generation of another example synthetic training dataset, in accordance with example embodiments.

[0034] FIG. 18 illustrates generation of another example synthetic training dataset, in accordance with example embodiments.Atty. Docket: 25-0667-WO

[0035] FIG. 19 is a diagram illustrating training and inference phases of a machine learning model, in accordance with example embodiments.

[0036] FIG. 20 depicts a distributed computing architecture, in accordance with example embodiments.

[0037] FIG. 21 is a block diagram of a computing device, in accordance with example embodiments.

[0038] FIG. 22 depicts a network of computing clusters arranged as a cloud-based server system, in accordance with example embodiments.

[0039] FIG. 23 is a flowchart of a method, in accordance with example embodiments.

[0040] FIG. 24 is another flowchart of a method, in accordance with example embodiments.DETAILED DESCRIPTIONOverview

[0041] This application generally relates to enabling efficient, real-time capture and digitization of multiple documents simultaneously using a mobile device’s camera. This approach addresses the limitations of existing scanning solutions, which are predominantly optimized for single-document capture and fail to adequately handle scenarios involving multiple documents within a single camera view without significant user friction or constraints. Also described herein is an approach to identify duplicate digital scans by performing a hierarchical comparison of respective visual features. The mobile document scanning approach described herein enables a seamless, intuitive, and intelligent multi-page capture experience. For the user, the experience is akin to capturing a video of multiple documents and the underlying system is configured to extract unique pages from the continuous video stream. Also described is a motion metering system to dynamically control frame capture timing.

[0042] Existing mobile document scanning solutions are generally configured to process one document at a time. Such detection algorithms are optimized to detect boundaries of a single document within the camera’s view. This is generally performed per capture cycle, which typically includes an interruptive preview step after each scan. Some existing approaches that allow multi-page scanning rely on rigid methods, such as segmenting the camera view to process specific regions independently (akin to splitting the screen for book scanning), rather than employing multi-object detection capable of identifying and isolating several documents arranged arbitrarily within a single, undivided frame. Existing multi -document methods allow for a maximum of two documents (e.g., two pages of an open book) to be scannedAtty. Docket: 25-0667-WOsimultaneously and enforce specific device orientations e.g., landscape mode) and precise document alignment. Also, such approaches split the screen into two halves and then process a single page found within the split screens. Such approaches cause the users to face a tedious workflow involving framing and scanning one page at a time before moving on to the next document.

[0043] Standard detection algorithms in existing approaches typically identify only one set of document boundaries per frame, failing to recognize or properly segment multiple distinct documents present simultaneously (e.g. , multiple receipts on a table, facing book pages, several business cards). This forces users into slow, sequential scanning. Existing auto-capture modes often rely on simple timers or basic stability checks. They lack the ability to recognize, during the capture session, if a document being viewed has already been scanned. This may result in user anxiety about accidentally creating duplicates and requires manual effort to identify and remove redundant pages later. It also prevents a truly fluid capture flow, as users must consciously avoid rescanning.

[0044] Such a single-document architecture limits performance and usability when multiple distinct documents need to be captured simultaneously from a single view. The singledocument architecture fails scanning in common scenarios, such as scanning multiple receipts laid out together, capturing facing pages of an open book, or digitizing several co-located small items (cards, photos, etc.). The central technical challenge is an inability of existing systems to perform real-time, simultaneous multi-document detection. Specifically, standard detection algorithms are not designed to identify and delineate more than one independent set of document boundaries within the same camera frame; they typically focus on a single detected item, effectively ignoring others present or just detecting it as one huge document.

[0045] Scanning is the process of collecting optical signals about physical objects and creating their digital representations hereby referred to as “assets” or “scans”. Digital scanning mechanisms often end up capturing multiple assets corresponding to the same object. For reducing compute load, it is often necessary to remove duplicate scans of the same object.

[0046] The digitization of physical documents, such as receipts, invoices, medical forms, notes, and so forth, is increasingly common for archival, processing, and / or data extraction purposes. Mobile device cameras offer a convenient means for such capture, allowing users to digitize documents anywhere. However, when capturing multiple documents in sequence, the introduction of duplicates is a significant practical problem.

[0047] Duplicate documents can arise for various reasons. For example, users might inadvertently scan the same entity twice, especially when dealing with large batches. CaptureAtty. Docket: 25-0667-WOprocesses might be interrupted and restarted, leading to rescanning of previously captured items. In semi-automated or video-based capture systems, where frames are analyzed continuously, slight movements or pauses by the user might cause the system to capture multiple visually identical or near-identical images of the same physical document before the user moves to the next one.

[0048] The presence of duplicate digital documents can result in several inefficiencies and potential errors. For example, the scanning system may consume unnecessary storage space, increasing the burden on subsequent processing steps (e.g., image enhancement, Optical Character Recognition (OCR), and / or data indexing), complicating document management, and can lead to user confusion or errors in analysis (e.g., double-counting expenses from duplicate receipts).

[0049] Existing deduplication approaches lack an ability to analyze document content during the capture phase of a multi-page session to intelligently prevent the autocapture of already-scanned pages. Instead, existing approaches rely on user input and / or postcapture cleanup. Since existing approaches are not configured with an intelligence to detect previously scanned documents, they often rely on triggering auto capture based on a fixed time interval and device stability signals. This can result in a slower scanning experience overall.

[0050] Users are generally required to manually review captured documents and delete duplicates, making the process time-consuming, tedious, and prone to human error, negating the benefits of rapid capture. Some systems perform deduplication after the entire capture session is complete. This requires storing all captured images initially, including duplicates, increasing temporary storage needs. The deduplication process itself can be computationally intensive and can delay the availability of the final, unique document set. Furthermore, such an approach does not provide real-time feedback to the user during capture.

[0051] Existing deduplication approaches that rely on cryptographic hashes (e.g., MD5, SHA-1) or pixel-based checksums can be error prone. They may fail to identify duplicates if there are small differences between two captures of the same document, such as slight variations in lighting, camera angle, scale, rotation, minor shadows, and / or compression artifacts, all of which are common in handheld video or rapid sequential photo capture.

[0052] Although comparing the extracted text content via OCR can identify duplicates for text-heavy documents, such an approach is often computationally expensive for real-time application during video capture. More importantly, it may perform sub-optimally and / or fail entirely for documents with little or no text (e.g., graphical receipts, diagrams, handwritten notes) or where OCR accuracy is compromised by image quality or non-standard fonts. TheAtty. Docket: 25-0667-WOOCR-based approach compares recognized content, not necessarily the document instance itself, thereby resulting in potential errors.

[0053] Techniques such as Scale-Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), or Oriented FAST and Rotated BRIEF (ORB) can match keypoints between images and offer more robustness than hashing. However, they may struggle with documents lacking sufficient distinct texture or features (e.g., plain forms). Furthermore, performing pairwise comparisons between a newly captured frame and potentially dozens or hundreds of previously captured documents in real-time during a video stream can become computationally prohibitive on resource-constrained devices like mobile devices.

[0054] Additionally, scans captured via mobile devices often suffer from imperfections like poor contrast, incorrect white balance, shadows from hands and / or lighting, and image noise. Existing approaches typically require users to manually apply filters or adjustments after capture to achieve a clean, professional look, adding extra steps to the workflow. Real-time correction during capture is generally absent.

[0055] Although post-capture filters are common in existing approaches, applying image enhancements (e.g., complex shadow removal, contrast / WB correction, denoising) in real-time during the capture preview or as part of the auto-capture decision process is not available. Users typically receive a capture with some corrections applied but often they have to manually remove shadows and perform other corrections later.

[0056] Another disadvantage of existing approaches is that they cannot identify and segment multiple, arbitrarily positioned documents within a single camera frame in real-time. Solutions attempting this are often restricted (e.g., to 2 pages) and rely on rigid layouts or screen segmentation, instead of dynamic object detection. For example, a “Book Scanning” mode may split the screen into two halves and then process a single page found within the split screens. Also, for example, capturing high-quality, sharp images of dynamic objects on resource-constrained mobile devices can be challenging due to motion blur.

[0057] From a perspective of a user, such limitations translate into wasted time and effort. For example, instead of efficiently capturing multiple receipts laid on a table or both pages of an open book in a single shot, users are tasked with a tedious, sequential process: frame one item, capture, wait for processing, then repeat for every single item individually. Further, the existing approaches are not configured with intelligent detection of unique documents in real time. Such approaches capture documents in the camera view based on a combination of time intervals and device stability, without identifying duplicate scans. This multi-step process negates the convenience mobile scanning aims to provide for such commonAtty. Docket: 25-0667-WOscenarios. The aforementioned technical challenges result in slower workflows, increased user effort and frustration, a potential for errors (e.g., duplicates), and inconsistent scan quality. The techniques described here provide technical solutions to the aforementioned technical challenges.

[0058] In summary, existing mobile document capture architectures are fundamentally limited by a single-object processing paradigm that fails to resolve the distinct technical hurdles of real-time, multi-document environments. These systems lack the requisite computational intelligence to perform simultaneous detection and segmentation of multiple, arbitrarily arranged documents within a single undivided frame, leading to inefficient resource utilization and suboptimal capture flows. Furthermore, traditional deduplication methods rely on computationally expensive post-capture analysis or error-prone pixel-based hashing that fails to account for the physical variability of handheld video capture, such as lighting fluctuations, perspective skews, and motion artifacts. Such limitations may result in a significant technical “bottleneck” where resource-constrained mobile devices must choose between real-time performance and scan accuracy. There exists, therefore, a technical need for a system that integrates machine-learning-based multi-object localization, hierarchical feature comparison for real-time duplicate prevention, and motion-adaptive capture timing. By transitioning from discrete, timer-based capture to an intelligent, continuous-stream analysis, the technical solutions described herein provide a concrete improvement to the computer's underlying functionality in processing high-fidelity digital assets while simultaneously reducing storage and processing overhead.Real-time multi-document capture

[0059] The technical solution described involves support for different camera modes. For example, a user need not be asked to use any particular mode landscape or portrait based on scanning document type. The technical solution described does not involve the user having to manually select a type of document to be scanned (e.g. , books, receipts, post-its, whiteboard, etc.). Documents may be scanned in any orientation, and in addition to single documents, more than two documents in a frame may be scanned. Also, users do not have to specify if they want to scan a single document or multiple documents. The model described herein is configured to automatically make such selections.

[0060] For example, an advanced computer vision processing pipeline and visionbased machine learning is described, typically operating on-device during a live camera preview. Unlike existing systems that detect one set of document boundaries, the technologiesAtty. Docket: 25-0667-WOdescribed herein perform real-time analysis of the camera feed. In some embodiments, a presence of multiple documents may be identified, and the contours of such multiple distinct documents may be delineated within a current frame (e.g., using machine learning techniques).

[0061] The techniques described herein may intelligently determine whether single or multiple documents are present and apply the appropriate detection logic without requiring user intervention and / or mode switching. Such techniques also allow for rapid, sequential, or simultaneous capture of detected documents without necessarily providing the user with a preview / confirmation screen after each individual item is captured, thereby enabling a fluid scanning session akin to taking a video.

[0062] For example, the system is capable of detecting and processing more than two distinct documents (e.g., potentially supporting up to 5, 10, 20, or more, subject to computational resources and visual complexity) present within a single camera frame at the same time. This overcomes the inability of existing scanning approaches to perform tasks like scanning multiple receipts spread on a table or capturing both facing pages of an open book in a single capture. This approach eliminates a need for separate, user-selected modes for single vs. multiple documents, or for specific document types (e.g., ‘book mode’, ‘receipt mode’, etc.).

[0063] In some embodiments, the underlying model may be configured to automatically adapt to various scenarios (e.g., single page, multiple pages, different document types like receipts, pages, notes, etc.) within a single, unified interface. This contrasts with existing approaches that often require explicit mode selection, adding complexity. The multidocument detection is configured to be independent of an orientation of a device (e.g., portrait or landscape) or an orientation of documents within the frame. This provides significantly more flexibility than existing approaches which may impose specific orientation requirements for multi-page features. By automatically detecting single or multiple documents and removing unnecessary workflow interruptions (such as forced previews after each capture), the approach described herein provides a streamlined and significantly faster process for digitizing multiple items compared to the cumbersome, one-at-a-time approach of existing scanners.

[0064] In some embodiments, a user may point their image capture device (e.g., camera) at one or more documents. The system may highlight the detected documents (one or multiple) in the live preview. Image capture may be automatically triggered in the event certain image quality criteria are satisfied for the detected documents, or the image capture may be user-initiated for the detected items (e.g., with a tap or automatically without a tap using an auto capture mode). For example, a user may capture two facing book pages simultaneously,Atty. Docket: 25-0667-WOand / or rapidly scan multiple receipts laid out on a desk in a single continuous operation without a need to individually frame and capture each one or switch modes.

[0065] A maximum number of documents that may be simultaneously scanned can be determined based on device capabilities. Also, for example, the multi -document detection logic may be integrated into operating system-level services and / or other applications such as file managers or cloud storage clients.

[0066] FIG. 1 is an example overview of a multi-document detector system 100, in accordance with example embodiments. An image capturing device 105 may include a viewfinder 110. As a user moves the image capturing device 105 over one or more documents, these documents appear in the viewfinder 110. A camera stream 115 can include a plurality of image frames which may be provided to a multi-page detector 120. Multi-page detector 120 may identify and capture (or scan) the one or more documents 125 in the camera stream 115. The scanned one or more documents 125 may be provided to an enhancement processor 130 that applies various image enhancements. In some embodiments, the enhancement processor 130 may include a machine learning (ML) model 135 that automatically applies the various image enhancements. In some embodiments, the enhanced versions of the scanned one or more documents 125 may be provided to an orientation correction model 140. Orientation correction model 140 may be configured to detect an orientation of each scanned document, and apply corrections to orient the scanned documents in a desirable manner. In some embodiments, a post-processing component 145 may apply post-processing operations to the scanned documents. As each document is progressively scanned, a scanned version appears in viewfinder 110. As described herein, the multi-document scanning, image enhancements, postprocessing, and other operations are performed in real-time.

[0067] FIG. 2 illustrates examples of multi-document detection 200, in accordance with example embodiments. Image 205 illustrates two pages of an open book. A first bounding box 210 is displayed (e.g., in viewfinder 110 of FIG. 1) to delineate the boundaries of first page Pl, and a second bounding box 215 is displayed (e.g., in viewfinder 110 of FIG. 1) to delineate the boundaries of second page P2. Image 220 illustrates several receipts in different orientations. Bounding boxes 225, 230, 235, and 240 delineate the boundaries of respective documents DI, D2, D3, and D4. As described herein, pages of a book, receipts, notes, whiteboard content, etc. may be scanned during one real-time scanning operation.

[0068] FIG. 3 illustrates an example machine learning (ML) model 300 for multidocument scanning, in accordance with example embodiments. In some embodiments, ML model 300 may be a neural network. Input image 305 may be a single frame of a plurality ofAtty. Docket: 25-0667-WOframes in a live preview. Input image 305 may be provided to a single shot multibox detector (SSD) 310. SSD 310 may include a single convolutional neural network to predict both the bounding boxes and object classes in input image 305. For example, SSD 310 may predict the bounding boxes and object classes illustrated in FIG. 2.

[0069] In some embodiments, ML model 300 may include separate models for bounding box detection and corner point detection. The bounding box detection may be performed by a document bounding box detector 320. In some embodiments, document bounding box detector 320 may include a MobileNet (e.g., MobileNet v2) and an SSD.MobileNet may include several layers, such as a Visual Geometry Group (VGG) model 315 (e.g., VGG- 16 or VGG- 19) that includes several convolution layers through Conv_5 with resolution 512. The output of VGG model 315 may be provided to a classifier layer (e.g., with dimensions 3x3x(4x(Classes+4))) which feeds into detection layer 330. In some embodiments, detection layer 330 may be configured to perform several thousands of detections per class.

[0070] The notation “3x3(4x(Classes+4))” generally means that the kernel size is 3x3 and the number of output channels or feature maps is 4x(Classes+4). The term “classes” generally refers to a number of object classes ML model 300 is trained to detect. In this context, the “classes” generally refer to different types of documents (e.g., receipts, pages, notes, etc.) or possibly just “document” vs. “non-document”. The “+4” accounts for four bounding box regression parameters. For each detected object, the model predicts a bounding box, which is typically defined by four parameters: the x and y coordinates of the top-left comer and the width and height of the box. The term “4x” generally refers to the number of anchor boxes (or prior boxes) at each location in a feature map. Anchor boxes may be pre-defined bounding boxes of different sizes and aspect ratios that enable ML model 300 to predict object locations more accurately. In this case, there are 4 different anchor boxes at each location. Accordingly, Therefore, the entire expression “3x3x(4x(Classes+4))” describes a classifier layer that performs a convolution operation with a 3x3 kernel and outputs a tensor with (4x(Classes+4)) channels, where each channel corresponds to either a class prediction or a bounding box regression parameter for one of the 4 anchor boxes.

[0071] The 4 anchor boxes at each location may be pre-configured with scales and aspect ratios optimized for standard document geometries. In some embodiments, the aspect ratios may be selected from a set comprising approximately 1:1.41 (A4 / ISO standard), 1:1.29 (US Letter), and 1:3 or greater (business receipts). By utilizing domain-specific anchor ratios, the ML model 300 can achieve faster convergence and higher intersection-over-union (loU) scores during the initial candidate document identification phase.Atty. Docket: 25-0667-WO

[0072] The output of VGG model 315 may also be fed into a convolution layer Conv6 (e.g., dimension 3x3x1024), whose output is fed into another convolution layer Conv7 (e.g., dimension 1x1x1024). As used herein, the notation “NxNxY” generally refers to a convolutional kernel (or filter) of sizeNxN, and Y denotes the number of output channels (also known as feature maps). This indicates that the kernel is an NxN pixel square and the layer outputs Y different feature maps. Accordingly, 3x3x1024 indicates that the kernel is a 3x3 pixel square and the layer outputs 1024 different feature maps.

[0073] In some embodiments, each of Conv 6 and Conv7 may be fully connected (FC) layers with 1024 output channels or neurons. So, the output of each of these layers is a vector of 1024 values. The output of Conv7 may be provided to a classifier layer (e.g., with dimensions 3x3x(6x(Classes+4))) (e.g., a classifier layer that performs a convolution operation with a 3x3 kernel and outputs a tensor with (6x(Classes+4)) channels, where each channel corresponds to either a class prediction or a bounding box regression parameter for one of the 6 anchor boxes) which feeds into detection layer 330. The architecture beyond Conv7 may include extra feature layers 325 including convolution layers Conv 8, Conv 9, Conv 10 and Conv 11.

[0074] For example, Conv8 may include a first layer (e.g., of dimension 1x1x256) and a second layer (e.g., of dimension 3x3x512-s2). The notation “s2” indicates that a stride of the convolution is 2 pixels, “si” indicates that a stride of the convolution is 1 pixel. The stride generally determines how many pixels the kernel is shifted over the input feature map at each step. A stride of 1 means the kernel moves one pixel at a time, and a stride of 2 means the kernel moves 2 pixels at a time.

[0075] In some embodiments, the output may feed into detection layer 330. Conv9 may include a first layer (e.g., of dimension 1x1x128) and a second layer (e.g., of dimension 3x3x256-s2). The output may feed into detection layer 330. ConvlO may include a first layer (e.g., of dimension 1x1x128) and a second layer (e.g., of dimension 3x3x256-sl). The output may feed into detection layer 330 as an output (e.g., of dimension 3x3x(4x(Classes+4))). Convll may include a first layer (e.g., of dimension 1x1x128) and a second layer (e.g., of dimension 3x3x256-sl). The output may feed into detection layer 330. The dimensions provided herein are for illustrative purposes alone and are non-limiting. Various architectural configurations with different numbers of layers, kernel sizes, strides, feature maps, and / or anchor boxes are possible.

[0076] In some embodiments, the output of detection layer 330 may be provided to a non-maximum suppression (NMS) component 335. NMS is a post-processing technique usedAtty. Docket: 25-0667-WOin object detection to eliminate duplicate detections and select the most relevant bounding boxes that correspond to detected objects. For example, in the event multiple bounding boxes are detected for the same object, then NMS component 335 can identify such occurrences and retain the bounding box with the highest confidence score, while suppressing the other bounding boxes that have a significant overlap with the retained bounding box.

[0077] The resulting identified documents 340 may be provided to classification and localization 405 of FIG. 4. In some embodiments, the output of the NMS component 335, comprising identified documents 340, undergoes a region-of-interest (ROI) normalization process before being provided to classification and localization 405. This process involves cropping the image frame to the coordinates of the detected bounding box and rescaling the cropped region to a fixed input resolution (e.g., 256 X 256 pixels) required by the precise corner localizer (LDR-Net). This normalization ensures that corner detection is scale-invariant and focused on the document textures, improving localization accuracy regardless of the document’s original distance from the image capturing device.

[0078] A precision value may be finetuned for detecting documents vs. rejecting nondocuments that have four comers. Also, the misclassification of non-document objects may be reduced. For example, while optimizing to minimize false positives, the model training may maintain a high recall rate so that actual documents are consistently detected. The segmentation model may be configured to accurately identify true and / or closest document corners even when fingers or folds partially obscure the document’s edges and / or corners. The system can be configured to be robust to common occlusions when users hold documents and / or there are stapled documents during capture. When documents are stacked or overlapping, the detection logic can reliably identify and target only the boundaries of the uppermost, fully visible document for scanning, ignoring partially visible edges or corners of underlying documents.

[0079] FIG. 4 illustrates an example of comer detection 400, in accordance with example embodiments. In some embodiments, the machine learning model includes a bounding box detector configured to detect a document boundary and a corner detector configured to detect document corners. For example, classification and localization 405 may include a comer points branch 410 to detect comer points, a borders branch 415 to detect borders and a classification branch 420 to determine whether a candidate document is a document to be scanned. In the context of object detection, localization generally refers to the task of identifying and determining the location of objects within an image with a high degree of accuracy, typically by drawing a bounding box around each object of interest. In some embodiments, corner localizer may be configured as a localized discrepancy representationAtty. Docket: 25-0667-WOnetwork (LDR-Net) that includes classification and localization 405, feature fusion 425 and backbone 430. The term “LDR-Net” as used herein, generally refers to a type of neural network architecture specifically designed for accurately locating corners in images, likely with an emphasis on document images or similar structured content.

[0080] In some embodiments, the neural network includes a MobileNet architecture with a multi-head attention module. The multi-head attention module may include custom filters, optimized for a particular use case within MobileNetV4. Attention mechanisms in neural networks allow the model to focus on specific parts of the input data that are most relevant, “multi-head” generally refers to having multiple attention mechanisms working in parallel, allowing the model to attend to different aspects of the input simultaneously. The term “custom filters” generally refers to specific filters designed or adjusted for the particular task of document comer detection. They help the model identify the features that are important for locating corners, such as edges, lines, and angles. For example, the module may be fine-tuned for the specific problem of multi-document scanning, configured to handle variations in document size, orientation, lighting, occlusions, backgrounds, textures, and so forth.

[0081] Feature fusion 425 combines the outputs of the corner points branch 410, borders branch 415, and classification branch 420 before moving on to post-processing 435. Backbone 430 may be configured to perform the feature extraction part of the precise comer localizer (LDRNet) CNN architecture. Backbone 430 includes the initial set of layers that process the input from FIG. 3 and extract the features that are used for the subsequent corner points branch 410, borders branch 415, and classification branch 420. The output of backbone 430 is provided for post-processing 435 and subsequently presented in the live viewfinder as output images 440.

[0082] FIG. 5 illustrates another example machine learning (ML) model 500 for multidocument scanning, in accordance with example embodiments. In some embodiments, the ML model 500 may be an integrated model. Some embodiments involve identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents. For example, as a user pans a camera, several objects may be detected (e.g., using saliency heat maps). Some objects may appear to be documents but may not be documents. Generally, document heatmaps may be predicted to identify the presence or absence of the plurality of candidate documents.

[0083] Some embodiments involve detecting, by ML model 500 and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document. For example, ML model 500 may be trained to infer corner-locationsAtty. Docket: 25-0667-WObased on the predicted document heatmaps. Also, for example, ML model 500 may be configured to infer the bounding boxes from the located corners.

[0084] Some embodiments involve selecting, by ML model 500 and based on the respective set of comers, one or more documents, of the plurality of candidate documents, that are to be digitally scanned. For example, once the respective set of comers are detected, ML model 500 may be trained using machine vision techniques to recognize whether the plurality of candidate documents include documents that are to be digitally scanned. For example, ML model 500 may be trained to recognize printed pages, receipts, notes, writing on a white board, chapter pages, text, images, etc. to identify and then select the one or more documents that are to be digitally scanned.

[0085] Some embodiments involve scanning, based on the respective set of corners, the one or more selected documents. For example, the corner locations may provide the boundaries of the selected one or more documents that are to be digitally scanned. The system may then capture the selected one or more documents one at a time without any further user input. For example, as the user pans the camera, the selected one or more documents may be scanned automatically. Some embodiments involve providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents. For example, as each document is scanned, they may be presented in a thumbnail version and stored sequentially.

[0086] For example, input image 505 may be input into one or more convolution layers 510 (e.g., C2-C6). Input image 505 may be, for example, a 640x640 image. In some embodiments, the neural network may include a MobileNet architecture. Such embodiments involve pruning one or more of the Universal Inverted Blocks of the MobileNet architecture to reduce model size and / or regularize the MobileNet model. In some embodiments, the neural network may include a combination of a MobileNet architecture, a Feature Pyramid Network (FPN) decoder with custom filters, a CenterNet head, and post-processing techniques.

[0087] An FPN generally refers to a feature extractor where the input is a single-scale image of an arbitrary size, and the outputs are proportionally sized feature maps at multiple levels. This may be performed in a fully convolutional fashion. CenterNet generally refers to an anchorless object detection architecture that replaces the classical Non-Maximum Suppression (e.g., NMS component 335) at the post process, with a more streamlined algorithm integrated into the Convolutional Neural Network (CNN) flow. This provides a significant advantage in terms of algorithmic elegance. For example, the confidence heat map may be usedAtty. Docket: 25-0667-WOdirectly to remove irrelevant predictions, without decoding bounding boxes and without breaking the deep CNN flow.

[0088] For example, the neural network may include a Feature Pyramid Network (FPN) or a BiFPN. Accordingly, one or more feature maps 515 may be extracted from different scales. For example, P2-P5 may be extracted respectively from C2-C5 of convolution layers 510. Also, feature map P6 may be extracted from C6 and may have respective dimensions (e.g., 10x10x256). The one or more feature maps 515 including P2-P5 may be concatenated and provided to context module 520. For example, P2 may have an associated dimension (e.g., 160x160x256), P3 may have an associated dimension (e.g., 80x80x256), P4 may have an associated dimension (e.g., 40x40x256), and P5 may have an associated dimension (e.g., 20x20x256). One instance of context module 520 is shown. In some embodiments, there may be multiple instances (e.g., x5) of context module 520. Context module 520 may include several layers that downsample the image (e.g., from 256 to 128 to 64), and each of the outputs may be concatenated to generate a fused output with an associated resolution (e.g., 256).

[0089] In some embodiments, training of ML model 500 may involve a multi-task loss 525. For example, multi-task loss 525 can include one or more of a bounding box loss, a classification loss, a comer point regression loss, a heatmap loss, or a keypoint offset prediction loss. ML model 500 may also be trained on a combination of all such losses. The term “bounding box loss” generally refers to a measure of how well a predicted bounding box aligns with the actual (or ground truth) bounding box. The bounding box loss quantifies the error in the predicted location, size, or shape of the bounding box around an object.

[0090] The term “classification loss” generally refers to a measure of how accurately the model is classifying whether or not a detected object is indeed a document (e.g., a receipt, page, or note) as opposed to something else. It quantifies the error in the model’s prediction of the object’s category or type. If the model identifies a document as a “document” and it actually is, then the loss would be low for this component. But if the model classifies a document as a “non-document” or misclassifies it as another type of object, the loss would be high, and this would guide the model’s learning during training.

[0091] The term “comer point regression loss” generally refers to a measure of how accurately the model is predicting the exact locations of the corners of a document or object within an image. It quantifies the error between the predicted comer coordinates and the actual (or ground truth) corner coordinates. This comer point regression loss may be used when the model is trained to directly predict the numerical coordinates of these corners. If the model predicts the corners to be very close to the true locations, the loss will be low; if the model’sAtty. Docket: 25-0667-WOpredictions are far off, the loss will be high, and this will influence how the model adjusts its parameters during training.

[0092] A heatmap loss is generally used in tasks where the model predicts the probability distribution of an object’s location or the locations of its keypoints. The network outputs a “heatmap,” which is a spatial representation where each pixel or location indicates the probability or confidence that a particular point (like a comer or a part of an object) is at that position. The term “heatmap loss” generally refers to a measure of how well the predicted heatmap matches the ground truth heatmap. The ground truth heatmap usually has high values (e.g., 1) at the correct locations and low values (e.g., 0) elsewhere. The loss function penalizes the model when its predicted heatmap has low values at true locations or high values at incorrect locations.

[0093] To convert the predicted heatmap into discrete numerical coordinates for scanning, the system may apply a sub-pixel maximum activation search. For example, a Soft-Argmax operation or a two-dimensional (2D) Gaussian fit may be performed on the local neighborhoods of the heatmap peaks to extract the final (x, y) coordinates of the document corners. This conversion allows the system to achieve sub-pixel precision, which can reduce perspective distortion during the subsequent dewarping and digital scanning stages.

[0094] The term “keypoints” generally refers to specific points of interest in an image, like corners, joints in a body, or specific features (e.g., facial features). Offset prediction means the model is learning to predict a small displacement (offset) from an initial or estimated location to the actual, precise location of the keypoint. Accordingly, the term “keypoint offset prediction loss” generally refers to a measure of how accurately the model is predicting these small adjustments to the keypoint locations. It quantifies the error between the predicted offsets and the actual offsets needed to reach the true keypoint positions. Such a loss function helps refine the model’s ability to pinpoint keypoints very precisely, especially when the initial keypoint estimations may be slightly off.

[0095] In some embodiments, each of the losses may be associated with a respective weight. Such embodiments involve tuning the respective weights to detect documents in one or more of a white-on-white scenario, a document on a document scenario, and a partial occlusion scenario. The output may be passed through a post-processing component 530 and subsequently presented in the live viewfinder as output images 535.

[0096] In some embodiments, knowledge distillation (e.g., online distillation) may be employed to transfer knowledge from a high-quality “teacher” model to a streamlined and enhanced “student” version. The student model is typically a smaller, more efficient model andAtty. Docket: 25-0667-WOis trained to mimic the behavior of a larger, more complex teacher model. The teacher, having higher capacity and likely better accuracy, transfers its knowledge to the student, enabling the student to perform better than if it were trained solely on the ground truth labels. Traditional knowledge distillation typically involves training the teacher first, then using it to train the student. Online distillation, however, involves training both the teacher and student models simultaneously. They interact and learn from each other during the training process. This can lead to faster convergence and more robust learning for the student.

[0097] In the context of document scanning, the teacher model may be a very deep CNN architecture or even a transformer-based model that is trained extensively on a large and diverse dataset of document images. This dataset can include various document types (e.g., invoices, receipts, forms, book pages, etc.), orientations, lighting conditions, backgrounds, textures, and / or image quality variations. The teacher model may be trained to accurately perform tasks including locating documents within an image, identifying a type of document (e.g., invoice vs. receipt), precisely identifying the corners of the document for perspective correction, extracting text from the document, and so forth.

[0098] The student model can be a smaller, more efficient student model. For example, this model can be lightweight, suitable for deployment on resource-constrained devices (e.g., mobile phones, embedded systems). In some embodiments, the student model may include a MobileNet-based architecture or other efficient CNN designs.

[0099] During training, both the teacher and student models may receive the same input document images. In addition to providing hard ground truth labels (e.g., “this is an invoice”), the teacher model may also provide “soft targets.” These soft targets may be probability distributions over the classes or locations, representing the teacher’s confidence in its predictions. For example, the teacher might say “80% probability it’s an invoice, 15% probability it’s a receipt, 5% probability it’s something else.” The student may be trained on a standard loss that measures accuracy relative to the ground truth labels. Also, the student model may be trained on a distillation loss that measures the difference between the student’s predictions and the teacher’s soft targets. Such a loss function can be calibrated to encourage the student to mimic the teacher’s nuanced understanding and confidence levels.[000100] Both the teacher and student models may be updated simultaneously during training. The student learns from the teacher’s soft targets, and the teacher itself continues to learn from the data and potentially from the student, especially in online distillation scenarios where models continuously adapt. The student model learns from the teacher model not justAtty. Docket: 25-0667-WOwhat the correct answer is, but also how the teacher arrives at that answer, including subtle patterns and relationships within the data.[000101] The knowledge transferred from the teacher model may enable the student model, despite its smaller size, to achieve significantly higher accuracy than it would if trained only on the ground truth. The streamlined student model can perform document scanning tasks efficiently on devices with limited computational resources. The student model can learn to be more robust to variations in document images (e.g., blur, noise, skew, etc.) by mimicking the teacher model’s ability to handle such variations. The student model can learn the specific nuances and patterns of different document types from the teacher, model leading to more accurate classification and extraction.[000102] In a real-world document scanning application, new types of documents or variations might emerge. Online distillation allows both the teacher model and student model to continuously adapt to this new data. The teacher model can also benefit from the student model, especially in situations where the student model might generalize or discover patterns that the teacher model initially might have missed.[000103] Training of ML models for corner detection may involve several hyperparameters. For example, an average keypoint confidence threshold may have a data type with a float range of [0,1], The average keypoint confidence threshold is an average confidence score for the vertex locations of a detected quadrilateral. Increasing this threshold can cause the system to detect quadrilaterals that have a very high level of confidence. The experience may become frustrating for high values because the system may fail to detect some legitimate documents. On the other hand, decreasing the threshold may increase the number of spurious detections. Accordingly, training the ML model involves tuning the average keypoint confidence threshold for optimal corner point detection.[000104] Another hyperparameter involves applying weights to each loss function for multi-objective training. This can also be a data type with a float range of [0,1], As described herein, the comer detection model may be trained by minimizing a loss function that incorporates multiple objectives such as image classification (doc vs. non-doc), document localization (bounding box location for each document in the image) and keypoint regression (coordinates of each comer for all the documents in the image). Each of these objectives may be associated with a weight indicative of a relative significance in the overall training objective. These weights may be tuned to achieve an appropriate balance between an accuracy with which the model is able to detect documents in complex corner case scenarios like white-on-white such as a white document in a white background, doc-on-doc such as a first document in theAtty. Docket: 25-0667-WObackground of a second document, and partial occlusion where one document partially hides portions of a second document.[000105] The ML model may be trained on multiple datasets. In some embodiments, each training dataset may be associated with a weight. This can also be a data type with a float range of [0,1], For example, the corner detection model may be trained on multiple datasets that represent different use cases. Some training datasets may include documents that may be scanned by humans. Some training datasets may include documents that are generated procedurally using a simulated environment. Each dataset may be associated with a weight that reflects how many samples are to be included from the dataset in a training batch. The higher the weight for a given dataset, the more the representation of the given dataset and the higher is an expected accuracy of the trained model on the type of samples in the given dataset.[000106] FIG. 6 illustrates an example multi-document scanning system 600, in accordance with example embodiments. Input image frames may be received as a viewfinder stream 605 and provided via a viewfinder of image capturing device 610. The viewfinder stream 605 generally refers to the camera stream coming from the camera stack that produces viewfinder frames. There are multiple options to handle this. This may typically use an approximately 2MP resolution like 1920x1080.[000107] A capture stream 615 generally refers to a high-resolution camera stream used for image processing and real time capture. This may typically use the highest resolution stream available in 16:9 or 4:3 aspect ratio (e.g., same as that of the viewfinder).[000108] Both viewfinder stream 605 and capture stream 615 may be processed by a graph-based framework for building multimodal (e.g., video, audio, and sensor) machine learning pipelines. Such a stack is generally used in different real time processing engines.[000109] In some embodiments, a motion metering system (not illustrated) may be used to detect local motion and skip frame processing unless the user stabilizes. For example, the capture stream 615 may be passed through detector 620. For example, detector 620 may be the ML model of FIGs. 3 and 4, or the ML model of FIG. 5. If no documents are detected, “no detection” 625, then the image frame is skipped. In the event one or more documents are detected, the frame is provided to a frame / page quality analysis component 635. The analysis of frame / page quality analysis component 635 may provide recommendations for autofocus 640. Also, for example, in the event a particular image frame has a quality below an image quality threshold, that image frame may be skipped at block 645.[000110] The motion metering system can be further configured to analyze motion across a temporal window to enhance capture reliability. For example, the system may evaluate a flowAtty. Docket: 25-0667-WOmagnitude within a defined window of consecutive image frames. This can improve performance over systems relying on the flow magnitude of a single current frame. Various statistical aggregations of this temporal data may be employed, including the determination of a minimum, maximum, or average flow magnitude across the window. In some embodiments, the system may utilize a maximum flow magnitude identified within the consecutive frames of the window to determine frame stability. By accounting for the peak motion within this temporal buffer, the system effectively filters out transient jitter or minor erratic movements of the image capturing device. This multi-frame analysis ensures that the capture of the object of interest is triggered when a sustained period of stability is detected, thereby preventing the acquisition of blurred textures during brief moments of instability that might otherwise be missed by a single-frame analysis.[000111] The detected documents may be provided to a crop and transform component 650 to crop the documents as needed. The output may be provided to a correct orientation component 655 that is configured to correct the orientations of the various detected documents. For example, correct orientation component 655 may involve ML based orientation detection and correction of document pages. Duplicate detection component 660 is configured to determine whether a scanned document is a duplicate of a previously scanned document. After performing deduplication, the unique set of scanned documents may be saved in a buffer 665.[000112] In some embodiments, the unique set of scanned documents may be rendered at block 670 to be provided via the viewfinder of image capturing device 610. The document scanning process may be a continuous process that begins when the user initiates it by pressing an image capture button. The user may also indicate a termination of the image capture session by pressing the image capture button. When the scanning session is terminated, an illumination correction component 675 may correct brightness levels, image intensity, and so forth. A postprocessing component 680 may perform additional corrections for orientation, document boundaries, etc. and a preview screen may present the scanned documents to the user.[000113] The document scanning process described herein may detect when there is more than one document in the viewfinder. It can identify the bounds and corners of each individual document page in the frame. After applying appropriate cropping, each document may be scanned individually, and each scanned document may be displayed as a separate document. Also, the user interface may be configured to provide a visual indication to a user that each page is scanned.Video To Scan DocumentAtty. Docket: 25-0667-WO[000114] As described with reference to FIG. 6, mobile document scanning can enable a seamless, intuitive, and intelligent multi-page capture experience. For the user, the experience can be akin to taking a video of their documents and the underlying system automatically extracts unique pages from the continuous stream. For example, the user may launch the document scanner via an entry-point (e.g., in ANDROID™, this can be via the CAMERA™ app, FILES™ app or DRIVE™ app. The document scanner may be configured to default to the continuous capture state that supports multi-page scanning. In some embodiments, motion metering logic may assess device stability using sensor data. Frames captured during excessive device motion, which are likely to be blurry, may be identified and potentially filtered out or flagged as low quality at this stage. Filtered frames may then be passed onto the downstream systems in the pipeline for further analysis.[000115] The system may continuously process the live camera preview feed frame-by-frame using on-device ML models described herein. For example, An ML model may analyze each frame to detect the presence and boundaries of one or more distinct documents simultaneously (e.g., by detector 620). The ML model can identify multiple documents arbitrarily placed within the frame (e.g., multiple receipts, open book pages, adjacent cards, etc.). For each potential document detected, another ML model may compare the content features against a cache of features from documents already captured (e.g., within the current scanning session). Such comparison may be made in real-time as the documents are scanned.[000116] In some embodiments, an image enhancement pipeline (e.g., blocks 635, 650, 655) processes the potential document region in real-time to correct contrast, white balance, shadows, image noise, etc., generating a preview of the enhanced scan. In the event that a detected document is of sufficient quality (based on metrics, including motion stability), and is determined not to be a duplicate of a previously captured document in the session (e.g. , at block 660), the system can automatically capture an enhanced version of that document region. Visual and / or haptic feedback may be provided to confirm the capture. Generally, the system may be configured to remain in capture mode without navigating to a preview screen. If multiple unique, high-quality documents are detected in a frame, they can potentially be captured simultaneously or in rapid succession.[000117] In the event a detected document is identified as a duplicate (e.g., at block 660), the system provides visual feedback in the live preview (indicating that the document has been previously scanned) and does not automatically trigger capture, preventing redundancy. The user can continue directing the camera at subsequent pages or documents (e.g, panning, turning pages of a multi-page document, etc.). The system repeats the aforementioned steps,Atty. Docket: 25-0667-WOaccumulating multiple unique page scans seamlessly (e.g., at block 665). For example, a numerical counter can display the number of pages that are captured. The user explicitly signals completion (e.g., taps a “Done” button). Only then are they taken to a preview / editing screen displaying captured pages in a consistent order (e.g., top-bottom, left-right based on capture position). In some embodiments, optional post-processing (e.g., crop adjustment, rotation, filter changes) may be applied (e.g., at block 680) before saving / sharing.[000118] The aforementioned simultaneous multi-document detection solves the limitation of single-document systems by allowing users to capture multiple co-located items (e.g., receipts, book pages, cards) efficiently in one go, resulting in significant reductions in time, effort, and computational resources. The real-time duplicate detection solves user anxiety about accidental rescans and eliminates manual duplicate document removal. The process enables a more fluid, confident auto-capture experience compared to time-based triggers, and makes auto-mode more reliable. The real-time auto-enhancement pipeline solves the problem of sub-optimal default scan quality. The process delivers crisp, clean, professional quality scans automatically by correcting common issues (e.g., contrast, WB, shadows, noise) during the capture phase, reducing the need for manual post-processing. The continuous capture flow solves the workflow friction of traditional scan-preview-return cycles. The process makes multi-page scanning intuitive and fast, like taking a video.[000119] As described herein, the enhancement pipeline may be implemented as a single end-to-end ML model instead of a cascade of individual correction models. Duplicate detection may utilize different feature extraction or comparison algorithms (e.g., perceptual hashing, Scale-Invariant Feature Transform (SIFT) features, etc.). Quality assessment for auto-capture can be configured to incorporate different and / or additional metrics. While implemented on-device, certain non-privacy-sensitive computations can potentially leverage server-side resources if connectivity is available. The user interface (UI) / user experience (UX) for providing feedback (e.g., capture confirmation, duplicate indication) could vary based on device settings, user settings, etc.[000120] FIG. 7A illustrates an example multi-document scanning process 700A, in accordance with example embodiments. Several instances of a graphical user interface (GUI) are illustrated as the scanning process proceeds. For example, GUI l displays a scanned version of a first document DI. GUI 2 illustrates how a user pans from the first document DI to a second document D2. A scaled down image (e.g., a thumbnail) 705 of the scanned first document appears in GUI 2. GUI 3 illustrates continued panning that brings the secondAtty. Docket: 25-0667-WOdocument D2 at or near the center of the preview frame. GUI 4 illustrates that the scanning of the second document D2 is initiated.[000121] GUI 5 illustrates a scanned version of the second document D2. GUI 6 illustrates how the user continues to pan from the second document D2 to the third document D3. Two thumbnails 710 of the first scanned document DI and the second scanned document D2 appear in GUI 6. GUI 7 illustrates continued panning that brings the third document D3 at or near the center of the preview frame. GUI 8 illustrates that the scanning of the third document D3 is initiated.[000122] FIG. 7B illustrates an additional multi-document scanning process 700B of FIG.7A, in accordance with example embodiments. GUI 9 illustrates a scanned version of the third document D3. Three thumbnails 715 of the first scanned document DI, the second scanned document D2, and the third scanned document D3 appear in GUI 10. GUI ll and GUI 12 show the effects of post-processing image filters being applied to the scanned document D3.[000123] GUI 13 illustrates a user interface for saving the scanned documents 720. The user may be provided with a selection of one or more file formats (e.g. , a toggle switch between .pdf and .jpg). A data entry field 725 may be provided to enter a name for the document folder to save the scanned documents 720. In some embodiments, a machine learning model e.g., a large language model) may analyze the scanned documents 720 and automatically suggest one or more recommended titles. GUI 14 illustrates a user interface where the scanned documents 735 are saved under the title “GroveryBills.pdf.”[000124] FIG. 8A illustrates another example of a multi-document scanning process 800A, in accordance with example embodiments. As the user pans the camera, a first document DI appears in the viewfinder of GUI 8A. GUI 8B shows that the first document DI may not be oriented properly. GUI 8C illustrates an orientation correction applied to the first document DI. GUI 8D illustrates that the scanning of the first document DI is initiated. GUI 8E illustrates that the scanning of the first document DI is completed. A thumbnail 805 of the scanned first document DI is displayed in GUI 8E.[000125] GUI 8F illustrates that the user may turn the page to scan the second document D2. For example, the first document DI and the second document D2 may be stapled together. GUI 8G illustrates how comer detection is applied to the second document D2. GUI 8H illustrates that the scanning of the second document D2 is initiated, and the stapled portion is accounted for.[000126] FIG. 8B further illustrates the example multi-document scanning process 800B of FIG. 8 A, in accordance with example embodiments. GUI 8I illustrates that the scanning ofAtty. Docket: 25-0667-WOthe second document D2 is completed. A thumbnail 810 of the scanned second document D2 is displayed in GUI 8J along with post-processing options. GUI 8K displays the scanned second document D2 subsequent to post-processing. GUI 8L illustrates a user interface for saving the scanned documents 815. The user may be provided with a selection of one or more file formats (e.g., a toggle switch between .pdf and .jpg). A data entry field may be provided to enter a name for the document folder to save the scanned documents 815. In some embodiments, a machine learning model (e.g., a large language model) may analyze the scanned documents 815 and automatically suggest one or more recommended titles. GUI 8M illustrates a user interface where the scanned documents 820 are saved.[000127] FIG. 9 illustrates an example multi-document scanning process 900, in accordance with example embodiments. GUI 9A illustrates first layout 905 wherein two pages of a book are open for scanning. GUI 9B illustrates an example of an iterative scanning process. For example, in second layout 910, the two pages of the book illustrated in GUI 9A may be numbered “1” and “2” and may be scanned in that order. As the user turns the pages of the book, additional pages may be sequentially numbered “3”, “4”, “5”, “6”, “7”, and “8” and subsequently scanned in that order. Additional and / or alternative numbering schemes may be configured for sequential scanning.[000128] Generally, a consistent ordering logic for documents detected within a single frame may be implemented. A default sequence may order documents based primarily on their vertical position (topmost first, determined by the top boundary’s Y-coordinate) and secondarily on their horizontal position (leftmost first, determined by the left boundary’s X-coordinate) to resolve ties for documents at similar vertical levels.[000129] Although machine learning based approaches for corner detection are described herein, other heuristic and / or non-alogorithmic approaches are possible as well. For example, a Harris Comer Detector can be used to analyze image regions for significant changes in intensity in both x and y directions. Comers may be identified where these changes are large in both directions, indicating a point where the image structure changes drastically. In some embodiments, a “cornerness” score may be determined based on image gradients. Another approach is a Shi-Tomasi Corner Detector that uses a different cornerness measure that tends to be more stable and provides better comer localization. It looks for a minimum eigenvalue of the structure tensor. As another example, a Features from Accelerated Segment Test (FAST) Corner Detector may be used to classify a pixel as a corner if a circle of pixels around it contains a certain number of contiguous pixels that are significantly brighter or darker than the center pixel. The FAST Comer Detector is generally designed for speed and may be used to detect anAtty. Docket: 25-0667-WOinitial set of corners and then combined with another algorithm to provide more refined results. In some embodiments, for specific types of documents, a template matching approach may be used where a predefined template of what a document corner looks like (e.g., a small image patch of a comer) can be compared to the document image and detect corners by identifying locations where the template best matches. In some embodiments, the edges of the document may be detected using edge detection (e.g, a machine vision approach based on, for example, a Canny edge detector). Then, line detection (e.g, Hough transform) may be used to find the lines that represent the document’s borders. The intersection points of these lines may be considered to be the corners.Post-processing Pipeline[000130] FIG. 10 illustrates an example unwarping 1000 of a scanned document, in accordance with example embodiments. For example, document 1005 is shown as warped. Unwarped version 1010 is illustrated.[000131] FIG. 11 illustrates example post-processing 1100 of a scanned document, in accordance with example embodiments. For example, image 1105 shows staple marks, image 1110 shows dog ears and image 1115 shows how a finger holding the document may appear in the preview. Image 1120 illustrates a scanned document where the staple marks, dog ears and fingers have been removed.[000132] FIG. 12 illustrates an example shadow removal 1200 of a scanned document, in accordance with example embodiments. Image 1205 illustrates a preview where a shadow (e.g. , of a user) appears over the document to be scanned. Image 1210 illustrates the scanned version without the shadow.[000133] The post-processing pipeline may be configured to facilitate various interactions of the user with the scanned documents. For example, the user may re-order the scanned documents, make corrections such as removing one or more scanned documents, removing duplicate documents that may not have been automatically identified as duplicates, adjusting the comer locations, orientation, document boundary, and so forth.Automated Identification of Duplicate Scanned Documents[000134] The approach described herein identifies duplicate digital scans (assets) by performing a hierarchical comparison of their respective visual features. Duplicate documents can arise for various reasons. For example, users might inadvertently scan the same entity twice, especially when dealing with large batches. Capture processes might be interrupted andAtty. Docket: 25-0667-WOrestarted, leading to rescanning of previously captured items. In semi-automated or video-based capture systems, where frames are analyzed continuously, slight movements or pauses by the user might cause the system to capture multiple visually identical or near-identical images of the same physical document before the user moves to the next one.[000135] Some embodiments involve determining whether a first document of the selected one or more documents is a duplicate of a second document of the scanned one or more documents, the first document having been identified subsequent to the second document. Such embodiments involve, upon a determination that the first document is a duplicate of the second document, providing, by the display screen of the image capturing device, a user-selectable option to scan the first document. For example, as a user pans the camera, a document may be selected and scanned. Subsequently, the same document may appear in the preview. The system may be configured to determine that the document in the preview is a duplicate of the previously scanned document. In such instances, the user may be prompted to indicate whether they intend to rescan that document.[000136] FIG. 13 illustrates an example document deduplication process 1300, in accordance with example embodiments. In some embodiments, the determining whether the first document is a duplicate of the second document involves determining one or more hierarchical feature-level differences between the first document and the second document. For example, both Scan 1 1302 and Scan 2 1304 may be independently processed through a series of feature extractors operating at different levels of abstraction.[000137] In some embodiments, Feature Extractor 1 1306 may extract low-level visual features, generating feature sets Al 1308 (from Scan 1 1302) and Bl 1310 (from Scan 2 1304). Low-level visual features refer to basic, fundamental characteristics of an image or video frame. These may be initial, raw pieces of information that a computer vision system extracts from the visual data. In the context of facial recognition, these features may include edges, corners, blobs, textures, colors, etc. For example, edges are boundaries or lines where there is a significant change in brightness or color. In a face, this might be the outline of the jawline, the edges of the eyes, or the lines of the lips. For a document, these may be the edges of the document. The comers may be points where two edges meet or intersect. Corners may provide key landmark information. For faces, comers around the eyes, mouth, and nose may be highly significant. For documents, the corners may define the contours and / or shape of the document. Blobs may be regions of similar intensity or color. These can represent areas like the cheeks, forehead, or other distinct parts of the face. Textures may be patterns or repeating structures within an image region. Skin texture, for example, is a type of low-level feature. Color mayAtty. Docket: 25-0667-WOinclude pixel color values (RGB, etc.). Color information can help distinguish different parts of the face or identify skin tones.[000138] In some embodiments, Feature Extractor 2 1314 may extract mid-level visual features, generating feature sets A2 1316 (from Scan 1 1302) and B2 1318 (from Scan 2 1304). Mid-level visual features sit between the raw, low-level features (like edges, corners, and colors) and the high-level semantic information (like “this is a document” or “this is a person”). Mid-level visual features represent more complex, structured patterns that are combinations or arrangements of low-level features. In the context of document scanning and general image processing, examples include textons, object parts, shapes, spatial relationships, repeating patterns, texture patterns, etc.[000139] Textons may be recurring micro-structures or “visual words” that capture local image patterns. They may represent specific text styles, line types, or decorative elements common in documents. Object parts may mean recognizing objects such as a paragraph block, a table, a signature line, a header, or a footer. These are components that make up the overall structure of a document. Shapes may include geometric shapes or contours of objects in the image, such as rectangular regions, lines, curves, and blobs of specific forms. For documents, shapes may include the overall rectangular shape of a page or the lines forming a table. Spatial relationships capture information about how different low-level features or parts of an object are arranged relative to each other. For example, knowing that a block of text is above a signature line is a mid-level spatial relationship. Repeating patterns include any recurring structures such as grids, lines in a table, or specific graphic elements found multiple times on a page. Texture patterns can include more complex textures beyond the basic textures. For example, distinguishing between the texture of printed text on paper versus the texture of a photograph included within a document may be a mid-level texture pattern.[000140] Mid-level features provide an abstraction layer between low-level data and high-level understanding. They make it easier for the system to transition from detecting basic details to recognizing meaningful objects or parts. In document scanning, understanding the layout and structure of the document may be highly significant. Mid-level features help to parse the document into sections, identify tables, and distinguish between text and images. Mid-level features may also provide context that low-level features generally cannot. For instance, an edge by itself is just an edge. But an edge that forms part of a rectangle might suggest the boundary of a text box or image within a document. Using mid-level features can make the system more robust to variations in lighting, noise, or slight distortions in the image. The structural and spatial relationships remain more consistent than the raw pixel values. ByAtty. Docket: 25-0667-WOcombining low-level features into meaningful patterns, mid-level features improve the accuracy of higher-level tasks like document classification, object detection within documents, and information extraction.[000141] In some embodiments, Feature Extractor 3 1322 may extract high-level visual features, generating feature sets A3 1324 (from Scan 1 1302) and B3 1326 (from Scan 2 1304). High-level visual features represent the semantic meaning and understanding of the content within an image. They represent features that are beyond basic shapes and textures to encompass complex concepts, objects, and their relationships. In the context of document scanning, high-level visual features may include document type recognition, structural layout understanding, key information extraction, semantic segmentation, document authenticity and / or validity, etc.[000142] For example, document type recognition may include identifying whether the document is an invoice, receipt, form, letter, or other specific type. As another example, structural layout understanding may include recognizing the logical organization of the document, such as the presence of headings, paragraphs, tables, signature blocks, etc. Key information extraction may involve isolating and understanding specific information such as dates, amounts, names, addresses, and other essential data points. Semantic segmentation may involve determining which parts of the image correspond to text, images, diagrams, or other elements. Document authenticity and / or validity may involve assessing if the document is genuine or fraudulent based on specific patterns, watermarks, seals, or other security features.[000143] High-level features enable the model to interpret the content and context of the document, rather than just its raw pixel data. They also enable automated document processing, reducing the need for manual review and data entry. High-level features facilitate efficient search and retrieval of documents based on their content and type. They enable data extraction and analysis, allowing insights to be gained from large volumes of scanned documents. They provide contextual information about the document, which is crucial for decision-making processes.[000144] In some embodiments, the differences between the corresponding feature sets from Scan 1 1302 and Scan 2 1304 may be calculated at each hierarchical level.[000145] For example, a Low-level difference: Al - Bl 1312, a Mid-level difference: A2 - B2 1320, and a High-level difference: A3 - B3 1328 may be determined and saved for further processing.[000146] FIG. 14 further illustrates the document deduplication process 1400 of FIG. 13, in accordance with example embodiments. In some embodiments, difference processing,Atty. Docket: 25-0667-WOfusion, summarization, and classification may be performed. Some embodiments involve fusing the one or more hierarchical feature-level differences to combine information across feature-levels. Such embodiments also involve summarizing the combined information across feature-levels. Such embodiments further involve, based on the summarized information, applying a classifier to determine whether the first document is a duplicate of the second document.[000147] For example, the input in the form of the stored hierarchical feature differences (Al - Bl, A2 - B2, A3 - B3) may be obtained from the process in FIG. 13. The differences from lower and mid-level features may undergo further processing.[000148] In some embodiments, the one or more hierarchical feature-level differences may include differences between low-level textures or edges that appear in the first document and the second document. For example, the low-level difference Al - Bl 1405 may be processed by a “Low-level difference processor” 1410 to yield a processed low-level difference representation, Diffl 1415.[000149] The mid-level difference A2 - B2 1420 may be processed by one or more “Midlevel difference processors” 1425 to yield processed mid-level difference representations, Diff2 1430 and Diff3 1435. Although not shown, A2-B2 may potentially feed into two processing paths or processors.[000150] In some embodiments, the one or more hierarchical feature-level differences may include differences between high-level object parts or shapes that appear in the first document and the second document. For example, the high-level difference A3 - B3 1440 may be generated.[000151] In some embodiments, feature fusion may be performed. For example, the processed low-level difference, Diffl 1415, the processed mid-level differences Diff2 1430, Diff3 1435, and the high-level difference A3 - B3 1440 may be combined in a “Feature-fusion” step 1445. This integrates the difference information across all hierarchical levels.[000152] In some embodiments, a summarization may be applied. For example, the fused feature representation may be passed through a “Summarization” module (not illustrated), to generate a concise representation of the overall difference between Scan 1 1302 and Scan 2 1304. The “Summarization” module may include a global average pooling (GAP) layer followed by a series of fully connected (FC) layers and a sigmoid activation function. The GAP layer reduces the spatial dimensions of the fused feature representation Diffl -3 and A3-B3 into a fixed-length feature vector. The subsequent FC layers process this vector to output a single scalar value between 0 and 1, representing the probability that Scan 1 and Scan 2 are duplicates.Atty. Docket: 25-0667-WOThis compact representation allows for real-time pairwise comparisons against a large cache of previously scanned assets without exceeding the memory or computational limits of the mobile device.[000153] The summarized difference representation may be used for classification. For example, a classification module may be configured to determine the final output, presumably indicating whether Scan 1 1302 and Scan 2 1304 are duplicates of the same object.[000154] An optional auxiliary task may be performed. For example, the summarized representation may be fed into an auxiliary task module, generating an auxiliary output for potentially related analyses or supplementary information.[000155] Such a hierarchical approach allows the system to compare scans at multiple levels of detail, from fine-grained low-level textures or edges to more abstract high-level object parts or shapes. Fusing the differences across these levels enables a robust determination of whether two scans represent the same physical object, thereby facilitating the removal of redundant data.[000156] The approach described herein is compatible with modern artificial neural network architectures like convolutional neural networks that are able to generalize across large collections of visual entities like images, videos, and three-dimensional (3D) point clouds. The approach is robust as it relies on the differences in visual features at varying levels of abstraction. Although three hierarchical levels are described, the process may be inductively modified to include more than three hierarchical levels to capture more fine grained details. Similarly, fewer than three hierarchical levels may be used. In some embodiments, a number of hierarchical levels may be based on a complexity of the document. For example, a document with easily discernible features (e.g., appearance of multiple fonts, font sizes, section headings, section numbers, distinctive visual layouts, etc.) may not require a lot of hierarchical levels.[000157] In some embodiments, the duplicate detection system may be configured as a multi-head architecture, comprising an embedding head and a classification head to facilitate robust feature representation and identity verification. The embedding head can include a 128-dimension projection layer designed to map visual features into a compact, latent space for efficient similarity comparison. During the training phase, the embedding head may be optimized using a Normalized Temperature-scaled Cross Entropy (NT-Xent) loss function, which serves as a modified contrastive loss to maximize agreement between differently augmented views of the same document while minimizing agreement with negative samples. To further refine the model’s discriminative power, the training process may employ onlineAtty. Docket: 25-0667-WOhard negative mining, which dynamically identifies and prioritizes challenging non-duplicate pairs that the model currently misclassifies as similar.[000158] In some embodiments, to ensure high-performance computational efficiency, the neural network and its associated operations may be implemented in an Accelerated Linear Algebra (XLA) compatible manner, enabling optimized execution on hardware accelerators such as Tensor Processing Units (TPUs). Additionally, the system may incorporate a variety of runtime augmentations, including the application of grayscale filters, rotations, and contrast shifts, to the training pipeline. These augmentations improve the generalizability of the duplicate detection model, ensuring it remains robust against variations in lighting, color, and device-specific sensor characteristics encountered during real-time mobile capture.[000159] Low level abstractions (from layers close to the input side) highlight differences in texture-details like lines, edges, color histograms. High level abstractions (from layers close to the output side) allow for differences in semantics such as part-whole relationships. This approach uses auxiliary objective functions along with the primary objective to classify duplicates. Auxiliary objectives can be used in supervised learning algorithms as a guidance mechanism to configure the processing modules to abide by the intended principles.[000160] The architecture described herein is designed with real time duplicate detection and removal in mind. Instead of entire scans, only important features of different scales can be cached for further processing. This prevents the accumulation of redundant information during the capture process making it feasible for execution on resource constrained devices like mobile phones.[000161] Such architecture avoids costly post-processing for offline duplicate removal. These algorithms usually involve heavy computation and are run on remote servers. In a mobile-phone based scanning application, the approach described herein reduces the amount of data communicated over the network by resolving all the basic cases on-device. Also, for example, real-time feedback and results are provided and a set of unique digital assets may be generated.[000162] The duplicate document detection model may be tuned to achieve a very low false positive rate. This ensures that new, unique documents that come into view during the scanning session are not incorrectly identified as duplicates and subsequently skipped by the auto-capture. For example, the ML model for document deduplication may involve one or more hyperparameters, such as, for example, a minimum confidence score. This can be a data type with a float range of [0,1], This is a lower-limit threshold on the predicted likelihood of a pair of document images being duplicated. Lowering this threshold can increase the number ofAtty. Docket: 25-0667-WOfalse-positives (e.g., non-duplicates predicted as duplicates). This case may be overcome using a manual override. Increasing the threshold may pass on duplicate pages downstream and increase workload of manual deduplication on the user. The model may be configured to handle blank documents and prevent auto-capture from continuously capturing blank documents.[000163] In some embodiments, the training of the ML model for document deduplication may utilize online hard negative mining within each training batch. The system may identify ‘hard negatives’ as pairs of different documents that currently yield high similarity embeddings, and may backpropagate the loss for these samples to force the model to learn subtle discriminative features, such as unique serial numbers or particular text layouts. This can be complemented by runtime augmentations including random grayscale conversion, which forces the model to rely on structural and textural patterns rather than color-based heuristics, thereby increasing generalizability across diverse document types and lighting conditions.[000164] FIG. 15 illustrates an example user interface for document deduplication 1500, in accordance with example embodiments. GUI 15A illustrates a camera panning over a document. GUI 15B illustrates how orientation correction is automatically applied to the document. The scanning process is illustrated in GUI 15C. GUI 15D illustrates the end of the scanning process. As the camera hovers over the document, the system determines in real-time that the document has been previously scanned. GUI 15E displays a message “Pages scanned” 1505. GUI 15F may optionally provide the user with a message “Tap to scan again” 1510. The user may tap the capture button to initiate another scan.Dynamic frame timing control for motion adaptive texture acquisition[000165] The approach described herein employs a motion metering system to dynamically control frame capture timing. This approach can be broadly applied to any image and / or video capture, and in particular, to document scanning. In the context of document scanning, an object of interest may be the document, and features may be the corners of the document. More generally, the features for an object of interest may be any significant features associated with the image and / or object of interest. For example, facial features may be key points for a face. Significant landmarks (e.g., monuments, mountains, building contours, etc.) may be key points for landmark photography.[000166] Although motion metering is described herein based on keypoints, the technique can be applied generally by tracking motion across image frames, instead of being applied toAtty. Docket: 25-0667-WOkeypoints. Keypoints may be used for latency minimization. However, motion may be measured in an image frame and used instead of keypoints.[000167] Some embodiments involve detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest. For example, keypoint detection may be performed by identifying a set of salient keypoints within each frame based on local shape and color pattern.[000168] Some embodiments involve tracking the one or more features in consecutive frames of the one or more image frames. For example, keypoint tracking may be performed by establishing correspondences between keypoints in consecutive frames using spatial context information.[000169] Some embodiments involve determining magnitudes of each motion vector corresponding to the tracked one or more features. For example, motion vectors may be determined by determining displacement vectors (motion vectors) of each tracked keypoint between consecutive frames. Some embodiments involve determining a motion magnitude by determining the magnitude of each motion vector (e.g., using Euclidean distance).[000170] In some embodiments, an aggregate motion metric (e.g., average, percentile such as P95, or P99) of the individual motion magnitudes may be determined.[000171] Some embodiments involve determining a frame stability parameter based on the aggregate motion metric. Generally speaking, frame stability is inversely proportional to the aggregate motion metric. In some embodiments, the frame stability parameter may be determined over a sliding temporal window of N consecutive frames (where N is typically between 5 and 15). The system may calculate a maximum motion magnitude observed across all tracked features within this window. By selecting the maximum (peak) motion rather than an average, the system ensures that capture is triggered if the device has remained consistently stable throughout the entire duration of the window, effectively suppressing accidental triggers caused by momentary pauses during a larger panning motion.[000172] Some embodiments involve triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest. For example, a threshold may be applied to the aggregate motion metric to perform binary classification to determine whether or not to capture the current frame.[000173] In the context of document scanning, a user may pan across several documents quickly without an intent to scan them. Also, for example, the user may scan a few pages of a book and quickly turn the pages to arrive at another set of pages that are to be scanned. Also, for example, the user may point the camera away from a document while arranging one or moreAtty. Docket: 25-0667-WOdocuments to be scanned. Several other instances may arise where documents may appear on a preview but are not intended for scanning. Accordingly, the system has to be configured to identify when the user intends for a document to be scanned, and trigger capture (or scanning) for such documents.[000174] For example, the scanning of the one or more selected documents involves, for each of the plurality of candidate documents, tracking the respective set of corners (e.g., as keypoints) in consecutive frames of the one or more image frames. Such embodiments involve determining whether the frame stability parameter exceeds a threshold stability value. Such embodiments also involve, upon a determination that the frame stability parameter exceeds the threshold stability value, triggering the scanning of the selected document.[000175] A classifier (e.g., a binary classifier) may be used for frame selection and may involve one or more hyperparameters, such as, for example, a minimum threshold value. This can be a data type with a float range of greater than or equal to 0. The optical flow (e.g., motion) magnitudes for multiple keypoint locations in the camera frame may be measured. In order to ensure that the selected frames are maximally stable, the 95-th percentile of keypoint motion magnitudes may be determined and the frame may be selected when the 95-th percentile is lower than the minimum threshold value. Increasing the minimum threshold value makes the system more lenient about local motion and some accepted frames may have motion blur. Some embodiments involve maintaining a low value for the threshold to allow high quality crisp frames to pass.[000176] Another hyperparameter may be a buffer size. This can be a data type with a float range of greater than or equal to 0. The buffer size represents the minimum number of stable frames that must be observed before comer detection is triggered. A smaller value would give a snappy user experience but may trigger comer detection on spurious frames captured if the camera happens to be momentarily stable during transit. A higher value may make the experience laggy as corner detection is expected to trigger the moment the user stabilizes the camera on a document.[000177] Motion metering enables precise and efficient motion estimation. An aggregate motion metric provides a robust measure of frame stability. The approach enables dynamic frame capture triggering based on real-time motion analysis and minimizes motion blur. The keypoint tracking and motion metering is designed to be computationally efficient enough to operate in real time on mobile devices.[000178] The approach reduces motion blur in captured images and minimizes memory consumption by capturing only stable frames. The approach may be optimized for real-timeAtty. Docket: 25-0667-WOperformance on mobile devices and results in improved accuracy in determining frame stability compared to sharpness-based methods. The approach is also less computationally expensive than frame synthesis.Training Datasets[000179] In some embodiments, the training of the machine learning model may be based on a synthetic training dataset. In some embodiments, the synthetic training dataset may have been generated prior to the training of the machine learning model. Several training datasets may be used for training the ML models described herein. For example, a data collection-multi-page (DC-MP) dataset (e.g., 5624 images) may include document images with labels including corners, visibility, and / or segmentation masks. A Blender Quad Simulated (BQS) dataset (e.g., 850,000 images) may include heavily augmented synthetic data for single documents. The BQS dataset is a dataset of synthetically generated images from Blender, a 3D modeling software. Virtual scanning environments may be created to capture images and ground truth labels using simulated cameras and ray tracing. The labeled data may include corners.[000180] In some embodiments, the synthetic training dataset may have been generated by using alpha blending to swap a background of an image of a captured document. For example, the synthetic background (BG) swap dataset (e.g., 10,000 images) may be synthetically generated by swapping backgrounds. The labels may include corners and / or segmentation masks.[000181] In some embodiments, the synthetic training dataset may have been generated by mixing images of two or more captured documents. For example, the synthetic mix match dataset (e.g., 20,000 images) may be synthetically generated by mixing cross images of documents. The labels may include corners and / or segmentation masks. A Mendeley dataset (e.g., 1403 images) may include a single corner dataset for corner case examples. The labeled data may include corners.[000182] FIG. 16 illustrates generation of an example synthetic training dataset, in accordance with example embodiments. Image 1600A illustrates document DI and document D2 over a black background BG1. Image 1600B illustrates document DI and document D2 over a textured background BG2. A large synthetic training dataset may be generated by switching the backgrounds for document DI and document D2 to capture different textures, patterns, colors, contrast levels, brightness levels, and so forth.[000183] FIG. 17 illustrates generation of another example synthetic training dataset, in accordance with example embodiments. For example, in addition to switching backgrounds,Atty. Docket: 25-0667-WOdifferent documents may be combined in various permutations and combinations. For example, documents DI, D2 and D3 are shown against a black background. Segmentation masks SD1, SD2, and SD3 corresponding to documents DI, D2 and D3, respectively, may be generated.[000184] Image 1705 illustrates how the segmentation masks SD1 and SD2 may be used to combine documents DI and D2 over a black boundary. Image 1710 illustrates how the segmentation masks SD2 and SD3 may be used to combine documents D2 and D3 over a black boundary.[000185] Image 1715 illustrates how image 1705 may be used to combine documents DI and D2 over a boundary BG1. Image 1720 illustrates how image 1710 may be used to combine documents D2 and D3 over another boundary BG2. As explained, various combinations of documents may be used against various backgrounds to generate a large synthetic training dataset.[000186] FIG. 18 illustrates generation of another example synthetic training dataset 1800, in accordance with example embodiments. As illustrated, using the processes described in FIGs. 16 and 17, combinations of documents over varied backgrounds BG1, BG2, BG3, and BG4 may be generated.[000187] Custom datasets for corner cases may be generated utilizing Blender, a 3D modeling platform with an integrated Physics Engine named Bullet. Cross-document mixing and matching may be performed using segmentation masks to produce an expanded multidocument dataset. The per corner confidence may be tuned along with the confidence of determining a presence or absence of a document.Training Machine Learning Models for Generating Inferences / Predictions [000188] FIG. 19 is a diagram 1900 illustrating a training phase 1902 and an inference phase 1904 of trained machine learning model(s) 1932, in accordance with example embodiments. Some machine learning techniques involve training one or more machine learning algorithms on an input set of training data to recognize patterns in the training data and provide output inferences and / or predictions about (patterns in the) training data. The resulting trained machine learning algorithm can be termed as a trained machine learning model. For example, FIG. 19 shows training phase 1902 where one or more machine learning algorithms 1920 are being trained on training data 1910 to become trained machine learning model(s) 1932. Then, during inference phase 1904, trained machine learning model(s) 1932 can receive input data 1930 and one or more inference / prediction requests 1940 (perhaps asAtty. Docket: 25-0667-WOpart of input data 1930) and responsively provide as an output one or more inferences and / or predict! on(s) 1950.[000189] As such, trained machine learning model(s) 1932 can include one or more models of one or more machine learning algorithms 1920. Machine learning algorithm(s) 1920 may include, but are not limited to: an artificial neural network (e.g., a herein-described convolutional neural networks, a recurrent neural network, a Bayesian network, a hidden Markov model, a Markov decision process, a logistic regression function, a support vector machine, a suitable statistical machine learning algorithm, and / or a heuristic machine learning system). Machine learning algorithm(s) 1920 may be supervised or unsupervised, and may implement any suitable combination of online and offline learning.[000190] In some examples, machine learning algorithm(s) 1920 and / or trained machine learning model(s) 1932 can be accelerated using on-device coprocessors, such as graphic processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), and / or application specific integrated circuits (ASICs). Such on-device coprocessors can be used to speed up machine learning algorithm(s) 820 and / or trained machine learning model(s) 1932. In some examples, trained machine learning model(s) 1932 can be trained, reside and execute to provide inferences on a particular computing device, and / or otherwise can make inferences for the particular computing device.[000191] During training phase 1902, machine learning algorithm(s) 1920 can be trained by providing at least training data 1910 as training input using unsupervised, supervised, semisupervised, and / or reinforcement learning techniques. Unsupervised learning involves providing a portion (or all) of training data 1910 to machine learning algorithm(s) 1920 and machine learning algorithm(s) 1920 determining one or more output inferences based on the provided portion (or all) of training data 1910. Supervised learning involves providing a portion of training data 1910 to machine learning algorithm(s) 1920, with machine learning algorithm(s) 1920 determining one or more output inferences based on the provided portion of training data 1910, and the output inference(s) are either accepted or corrected based on correct results associated with training data 1910. In some examples, supervised learning of machine learning algorithm(s) 1920 can be governed by a set of rules and / or a set of labels for the training input, and the set of rules and / or set of labels may be used to correct inferences of machine learning algorithm(s) 1920.[000192] Semi-supervised learning involves having correct results for part, but not all, of training data 1910. During semi-supervised learning, supervised learning is used for a portion of training data 1910 having correct results, and unsupervised learning is used for a portion ofAtty. Docket: 25-0667-WOtraining data 1910 not having correct results. Reinforcement learning involves machine learning algorithm(s) 1920 receiving a reward signal regarding a prior inference, where the reward signal can be a numerical value. During reinforcement learning, machine learning algorithm(s) 1920 can output an inference and receive a reward signal in response, where machine learning algorithm(s) 1920 are configured to try to maximize the numerical value of the reward signal. In some examples, reinforcement learning also utilizes a value function that provides a numerical value representing an expected total of the numerical values provided by the reward signal over time. In some examples, machine learning algorithm(s) 1920 and / or trained machine learning model(s) 1932 can be trained using other machine learning techniques, including but not limited to, incremental learning and curriculum learning.[000193] In some examples, machine learning algorithm(s) 1920 and / or trained machine learning model(s) 1932 can use transfer learning techniques. For example, transfer learning techniques can involve trained machine learning model(s) 1932 being pre-trained on one set of data and additionally trained using training data 1910. More particularly, machine learning algorithm(s) 1920 can be pre-trained on data from one or more computing devices and a resulting trained machine learning model provided to computing device CD1, where CD1 is intended to execute the trained machine learning model during inference phase 1904. Then, during training phase 1902, the pre-trained machine learning model can be additionally trained using training data 1910, where training data 1910 can be derived from kernel and non-kernel data of computing device CD1. This further training of the machine learning algorithm(s) 1920 and / or the pre-trained machine learning model using training data 1910 of CDl’s data can be performed using either supervised or unsupervised learning. Once machine learning algorithm(s) 1920 and / or the pre-trained machine learning model has been trained on at least training data 1910, training phase 1902 can be completed. The trained resulting machine learning model can be utilized as at least one of trained machine learning model(s) 1932.[000194] In particular, once training phase 1902 has been completed, trained machine learning model(s) 1932 can be provided to a computing device, if not already on the computing device. Inference phase 1904 can begin after trained machine learning model(s) 1932 are provided to computing device CD1.[000195] During inference phase 1904, trained machine learning model(s) 1932 can receive input data 1930 and generate and output one or more corresponding inferences and / or prediction(s) 1950 about input data 1930. As such, input data 1930 can be used as an input to trained machine learning model(s) 1932 for providing corresponding inference(s) and / or prediction(s) 1950 to kernel components and non-kernel components. For example, trainedAtty. Docket: 25-0667-WOmachine learning model(s) 1932 can generate inference(s) and / or predict! on(s) 1950 in response to one or more inference / prediction requests 1940. In some examples, trained machine learning model(s) 1932 can be executed by a portion of other software. For example, trained machine learning model(s) 1932 can be executed by an inference or prediction daemon to be readily available to provide inferences and / or predictions upon request. Input data 1930 can include data from computing device CD1 executing trained machine learning model(s) 1932 and / or input data from one or more computing devices other than CD1. For example, input data 1930 can include labeled or unlabeled input data items.[000196] Inference(s) and / or prediction(s) 1950 can include output at least one suggested entry, numerical values, and / or other output data (e.g., rewards, scores, etc.) produced by trained machine learning model(s) 1932 operating on input data 1930 (and training data 1910). In some examples, trained machine learning model(s) 1932 can use output inference(s) and / or predict! on(s) 1950 as input feedback 1960. Trained machine learning model(s) 1932 can also rely on past inferences as inputs for generating new inferences.[000197] A neural network such as ML models described herein can be examples of machine learning algorithm(s) 1920. After training, the trained version of the neural network can be an example of trained machine learning model(s) 1932, such as, for example, a trained ML model. In this approach, an example of the one or more inference / prediction request(s) 1940 can be a request to scan a document and a corresponding example of inferences and / or prediction(s) 1950 can be a predicted scanned document.[000198] In some examples, one computing device CD SOLO can include the trained version of the neural network, perhaps after training. Then, computing device CD SOLO can receive a request to detect a document, and use the trained version of the neural network to predict the presence of the document.[000199] In some examples, two or more computing devices CD CLI and CD SRV can be used to detect the document; e.g., a first computing device CD CLI can generate and send requests to detect a document to a second computing device CD SRV. Then, CD SRV can use the trained version of the neural network, to detect the document, and respond to the requests from CD CLI for the detected document. Then, upon reception of responses to the requests, CD CLI can provide the detected document (e.g., using a user interface and / or a display).Example Data Network[000200] FIG. 20 depicts a distributed computing architecture 2000, in accordance with example embodiments. Distributed computing architecture 2000 includes server devices 2008,Atty. Docket: 25-0667-WO2010 that are configured to communicate, via network 2006, with programmable devices 2004a, 2004b, 2004c, 2004d, 2004e. Network 2006 may correspond to a local area network (LAN), a wide area network (WAN), a WLAN, a WWAN, a corporate intranet, the public Internet, or any other type of network configured to provide a communications path between networked computing devices. Network 2006 may also correspond to a combination of one or more LANs, WANs, corporate intranets, and / or the public Internet.[000201] Although FIG. 20 only shows five programmable devices, distributed application architectures may serve tens, hundreds, or thousands of programmable devices. Moreover, programmable devices 2004a, 2004b, 2004c, 2004d, 2004e (or any additional programmable devices) may be any sort of computing device, such as a mobile computing device, desktop computer, wearable computing device, head-mountable device (HMD), network terminal, a mobile computing device, and so on. In some examples, such as illustrated by programmable devices 2004a, 2004b, 2004c, 2004e, programmable devices can be directly connected to network 2006. In other examples, such as illustrated by programmable device 2004d, programmable devices can be indirectly connected to network 2006 via an associated computing device, such as programmable device 2004c. In this example, programmable device 2004c can act as an associated computing device to pass electronic communications between programmable device 2004d and network 2006. In other examples, such as illustrated by programmable device 2004e, a computing device can be part of and / or inside a vehicle, such as a car, a truck, a bus, a boat or ship, an airplane, etc. In other examples not shown in FIG. 20, a programmable device can be both directly and indirectly connected to network 2006.[000202] Server devices 2008, 2010 can be configured to perform one or more services, as requested by programmable devices 2004a-2004e. For example, server device 2008 and / or 2010 can provide content to programmable devices 2004a-2004e. The content can include, but is not limited to, web pages, hypertext, scripts, binary data such as compiled software, images, audio, and / or video. The content can include compressed and / or uncompressed content. The content can be encrypted and / or unencrypted. Other types of content are possible as well.[000203] As another example, server device 2008 and / or 2010 can provide programmable devices 2004a-2004e with access to software for database, search, computation, graphical, audio, video, World Wide Web / Internet utilization, and / or other functions. Many other examples of server devices are possible as well.Computing Device ArchitectureAtty. Docket: 25-0667-WO[000204] FIG. 21 is a block diagram of a computing device 2100, in accordance with example embodiments. In particular, computing device 2100 shown in FIG. 21 can be configured to perform at least one function of and / or related to the machine learning models, method 2300, and / or method 2400.[000205] Computing device 2100 may include a user interface module 2101, a network communications module 2102, one or more processors 2103, data storage 2104, one or more camera(s) 2112, one or more sensors 2114, and power system 2116, all of which may be linked together via a system bus, network, or other connection mechanism 2105.[000206] User interface module 2101 can be operable to send data to and / or receive data from external user input / output devices. For example, user interface module 2101 can be configured to send and / or receive data to and / or from user input devices such as a touch screen, a computer mouse, a keyboard, a keypad, a touch pad, a trackball, a joystick, a voice recognition module, and / or other similar devices. User interface module 2101 can also be configured to provide output to user display devices, such as one or more cathode ray tubes (CRT), liquid crystal displays, light emitting diodes (LEDs), displays using digital light processing (DLP) technology, printers, light bulbs, and / or other similar devices, either now known or later developed. User interface module 2101 can also be configured to generate audible outputs, with devices such as a speaker, speaker jack, audio output port, audio output device, earphones, and / or other similar devices. User interface module 2101 can further be configured with one or more haptic devices that can generate haptic outputs, such as vibrations and / or other outputs detectable by touch and / or physical contact with computing device 2100. In some examples, user interface module 2101 can be used to provide a graphical user interface (GUI) for utilizing computing device 2100, such as, for example, a graphical user interface of a mobile phone device.[000207] Network communications module 2102 can include one or more devices that provide one or more wireless interface(s) 2107 and / or one or more wireline interface(s) 2108 that are configurable to communicate via a network. Wireless interface(s) 2107 can include one or more wireless transmitters, receivers, and / or transceivers, such as a Bluetooth™ transceiver, a Zigbee® transceiver, a Wi-Fi™ transceiver, a WiMAX™ transceiver, an LTE™ transceiver, and / or other type of wireless transceiver configurable to communicate via a wireless network. Wireline interface(s) 2108 can include one or more wireline transmitters, receivers, and / or transceivers, such as an Ethernet transceiver, a Universal Serial Bus (USB) transceiver, or similar transceiver configurable to communicate via a twisted pair wire, a coaxial cable, a fiber-optic link, or a similar physical connection to a wireline network.Atty. Docket: 25-0667-WO[000208] In some examples, network communications module 2102 can be configured to provide reliable, secured, and / or authenticated communications. For each communication described herein, information for facilitating reliable communications (e.g., guaranteed message delivery) can be provided, perhaps as part of a message header and / or footer (e.g., packet / message sequencing information, encapsulation headers and / or footers, size / time information, and transmission verification information such as cyclic redundancy check (CRC) and / or parity check values). Communications can be made secure (e.g., be encoded or encrypted) and / or decry pted / decoded using one or more cryptographic protocols and / or algorithms, such as, but not limited to, Data Encryption Standard (DES), Advanced Encryption Standard (AES), a Rivest- Shamir- Adelman (RSA) algorithm, a Diffie-Hellman algorithm, a secure sockets protocol such as Secure Sockets Layer (SSL) or Transport Layer Security (TLS), and / or Digital Signature Algorithm (DSA). Other cryptographic protocols and / or algorithms can be used as well or in addition to those listed herein to secure (and then decry pt / decode) communications.[000209] One or more processors 2103 can include one or more general purpose processors, and / or one or more special purpose processors (e.g., digital signal processors, tensor processing units (TPUs), graphics processing units (GPUs), application specific integrated circuits, etc.). One or more processors 2103 can be configured to execute computer-readable instructions 2106 that are contained in data storage 2104 and / or other instructions as described herein.[000210] Data storage 2104 can include one or more non-transitory computer-readable storage media that can be read and / or accessed by at least one of one or more processors 2103. The one or more computer-readable storage media can include volatile and / or non-volatile storage components, such as optical, magnetic, organic or other memory or disc storage, which can be integrated in whole or in part with at least one of one or more processors 2103. In some examples, data storage 2104 can be implemented using a single physical device (e.g., one optical, magnetic, organic or other memory or disc storage unit), while in other examples, data storage 2104 can be implemented using two or more physical devices.[000211] Data storage 2104 can include computer-readable instructions 2106 and perhaps additional data. In some examples, data storage 2104 can include storage required to perform at least part of the herein-described methods, scenarios, and techniques and / or at least part of the functionality of the herein-described devices and networks. In some examples, data storage 2104 can include storage for a trained neural network model 2110 (e.g., trained machine learning model, etc.). In particular of these examples, computer-readable instructions 2106 canAtty. Docket: 25-0667-WOinclude instructions that, when executed by one or more processors 2103, enable computing device 2100 to provide for some or all of the functionality of trained neural network model 2110.[000212] In one aspect, the functions include: identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents; detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of comers indicative of a boundary of each candidate document; selecting, by the machine learning model and based on the respective set of corners, one or more documents, of the plurality of candidate documents, that are to be digitally scanned; scanning, based on the respective set of corners, the one or more selected documents; and providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.[000213] In another aspect, the functions include: detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest; tracking the one or more features in consecutive frames of the one or more image frames; determining magnitudes of each motion vector corresponding to the tracked one or more features; determining an aggregate motion metric for the magnitudes of each motion vector; determining a frame stability parameter based on the aggregate motion metric; and triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.[000214] In some examples, computing device 2100 can include one or more camera(s) 2112. Camera(s) 2112 can include one or more image capture devices, such as still and / or video cameras, equipped to capture light and record the captured light in one or more images; that is, camera(s) 2112 can generate image(s) of captured light. The one or more images can be one or more still images and / or one or more images utilized in video imagery. Camera(s) 2112 can capture light and / or electromagnetic radiation emitted as visible light, infrared radiation, ultraviolet light, and / or as one or more other frequencies of light.[000215] In some examples, computing device 2100 can include one or more sensors 2114. Sensors 2114 can be configured to measure conditions within computing device 2100 and / or conditions in an environment of computing device 2100 and provide data about these conditions. For example, sensors 2114 can include one or more of: (i) sensors for obtaining data about computing device 2100, such as, but not limited to, a thermometer for measuring a temperature of computing device 2100, a battery sensor for measuring power of one or more batteries of power system 2116, and / or other sensors measuring conditions of computing deviceAtty. Docket: 25-0667-WO2100; (ii) an identification sensor to identify other objects and / or devices, such as, but not limited to, a Radio Frequency Identification (RFID) reader, proximity sensor, one-dimensional barcode reader, two-dimensional barcode (e.g., Quick Response (QR) code) reader, and a laser tracker, where the identification sensors can be configured to read identifiers, such as RFID tags, barcodes, QR codes, and / or other devices and / or object configured to be read and provide at least identifying information; (iii) sensors to measure locations and / or movements of computing device 2100, such as, but not limited to, a tilt sensor, a gyroscope, an accelerometer, a Doppler sensor, a GPS device, a sonar sensor, a radar device, a laser-displacement sensor, and a compass; (iv) an environmental sensor to obtain data indicative of an environment of computing device 2100, such as, but not limited to, an infrared sensor, an optical sensor, a light sensor, a biosensor, a capacitive sensor, a touch sensor, a temperature sensor, a wireless sensor, a radio sensor, a movement sensor, a microphone, a sound sensor, an ultrasound sensor and / or a smoke sensor; and / or (v) a force sensor to measure one or more forces (e.g., inertial forces and / or G-forces) acting about computing device 2100, such as, but not limited to one or more sensors that measure: forces in one or more dimensions, torque, ground force, friction, and / or a zero moment point (ZMP) sensor that identifies ZMPs and / or locations of the ZMPs. Many other examples of sensors 2114 are possible as well.[000216] Power system 2116 can include one or more batteries 2118 and / or one or more external power interfaces 2120 for providing electrical power to computing device 2100. Each battery of the one or more batteries 2118 can, when electrically coupled to the computing device 2100, act as a source of stored electrical power for computing device 2100. One or more batteries 2118 of power system 2116 can be configured to be portable. Some or all of one or more batteries 2118 can be readily removable from computing device 2100. In other examples, some or all of one or more batteries 2118 can be internal to computing device 2100, and so may not be readily removable from computing device 2100. Some or all of one or more batteries 2118 can be rechargeable. For example, a rechargeable battery can be recharged via a wired connection between the battery and another power supply, such as by one or more power supplies that are external to computing device 2100 and connected to computing device 2100 via the one or more external power interfaces. In other examples, some or all of one or more batteries 2118 can be non-rechargeable batteries.[000217] One or more external power interfaces 2120 of power system 2116 can include one or more wired-power interfaces, such as a USB cable and / or a power cord, that enable wired electrical power connections to one or more power supplies that are external to computing device 2100. One or more external power interfaces 2120 can include one or moreAtty. Docket: 25-0667-WOwireless power interfaces, such as a Qi wireless charger, that enable wireless electrical power connections, such as via a Qi wireless charger, to one or more external power supplies. Once an electrical power connection is established to an external power source using one or more external power interfaces 820, computing device 2100 can draw electrical power from the external power source the established electrical power connection. In some examples, power system 2116 can include related sensors, such as battery sensors associated with the one or more batteries or other types of electrical power sensors.Cloud-Based Servers[000218] FIG. 22 depicts a network of computing clusters arranged as a cloud-based server system, in accordance with example embodiments. In FIG. 22, functionality of a neural network, and / or a computing device can be distributed among computing clusters 2209a, 2209b, and 2209c. Computing cluster 2209a can include one or more computing devices 2200a, cluster storage arrays 2210a, and cluster routers 2211a connected by a local cluster network 2212a. Similarly, computing cluster 2209b can include one or more computing devices 2200b, cluster storage arrays 2210b, and cluster routers 2211b connected by a local cluster network 2212b. Likewise, computing cluster 2209c can include one or more computing devices 2200c, cluster storage arrays 2210c, and cluster routers 2211c connected by a local cluster network 2212c.[000219] In some embodiments, computing clusters 2209a, 2209b, and 2209c can be a single computing device residing in a single computing center. In other embodiments, computing clusters 2209a, 2209b, and 2209c can include multiple computing devices in a single computing center, or even multiple computing devices located in multiple computing centers located in diverse geographic locations. For example, FIG. 22 depicts each of computing clusters 2209a, 2209b, 2209c residing in different physical locations.[000220] In some embodiments, data and services at computing clusters 2209a, 2209b, 2209c can be encoded as computer readable information stored in non-transitory, tangible computer readable media (or computer readable storage media) and accessible by other computing devices. In some embodiments, computing clusters 2209a, 2209b, 2209c can be stored on a single disk drive or other tangible storage media, or can be implemented on multiple disk drives or other tangible storage media located at one or more diverse geographic locations.[000221] In some embodiments, each of computing clusters 2209a, 2209b, and 2209c can have an equal number of computing devices, an equal number of cluster storage arrays, and an equal number of cluster routers. In other embodiments, however, each computing cluster canAtty. Docket: 25-0667-WOhave different numbers of computing devices, different numbers of cluster storage arrays, and different numbers of cluster routers. The number of computing devices, cluster storage arrays, and cluster routers in each computing cluster can depend on the computing task or tasks assigned to each computing cluster.[000222] In computing cluster 2209a, for example, computing devices 2200a can be configured to perform various computing tasks of a neural network, and / or a computing device. In one embodiment, the various functionalities of a neural network, and / or a computing device can be distributed among one or more of computing devices 2200a, 2200b, and 2200c. Computing devices 2200b and 2200c in respective computing clusters 2209b and 2209c can be configured similarly to computing devices 2200a in computing cluster 2209a. On the other hand, in some embodiments, computing devices 2200a, 2200b, and 2200c can be configured to perform different functions.[000223] In some embodiments, computing tasks and stored data associated with a neural network, and / or a computing device can be distributed across computing devices 2200a, 2200b, and 2200c based at least in part on the processing requirements of a neural network, and / or a computing device, the processing capabilities of computing devices 2200a, 2200b, 2200c, the latency of the network links between the computing devices in each computing cluster and between the computing clusters themselves, and / or other factors that can contribute to the cost, speed, fault-tolerance, resiliency, efficiency, and / or other design goals of the overall system architecture.[000224] Cluster storage arrays 2210a, 2210b, 2210c of computing clusters 2209a, 2209b, and 2209c can be data storage arrays that include disk array controllers configured to manage read and write access to groups of hard disk drives. The disk array controllers, alone or in conjunction with their respective computing devices, can also be configured to manage backup or redundant copies of the data stored in the cluster storage arrays to protect against disk drive or other cluster storage array failures and / or network failures that prevent one or more computing devices from accessing one or more cluster storage arrays.[000225] Similar to the manner in which the functions of a conditioned, axial selfattention based neural network, and / or a computing device can be distributed across computing devices 2200a, 2200b, 2200c of computing clusters 2209a, 2209b, 2209c, various active portions and / or backup portions of these components can be distributed across cluster storage arrays 2210a, 2210b, 2210c. For example, some cluster storage arrays can be configured to store one portion of the data of a first layer of a neural network, and / or a computing device, while other cluster storage arrays can store other portion(s) of data of second layer of a neuralAtty. Docket: 25-0667-WOnetwork, and / or a computing device. Also, for example, some cluster storage arrays can be configured to store the data of an encoder of a neural network, while other cluster storage arrays can store the data of a decoder of a neural network. Additionally, some cluster storage arrays can be configured to store backup versions of data stored in other cluster storage arrays.[000226] Cluster routers 2211a, 2211b, 2211c in computing clusters 2209a, 2209b, and 2209c can include networking equipment configured to provide internal and external communications for the computing clusters. For example, cluster routers 221 la in computing cluster 2209a can include one or more internet switching and routing devices configured to provide (i) local area network communications between computing devices 2200a and cluster storage arrays 2210a via local cluster network 2212a, and (ii) wide area network communications between computing cluster 2209a and computing clusters 2209b and 2209c via wide area network link 2213a to network 2006. Cluster routers 2211b and 2211c can include network equipment similar to cluster routers 2211a, and cluster routers 2211b and 2211c can perform similar networking functions for computing clusters 2209b and 2209b that cluster routers 2211a perform for computing cluster 2209a.[000227] In some embodiments, the configuration of cluster routers 2211a, 2211b, 2211c can be based at least in part on the data communication requirements of the computing devices and cluster storage arrays, the data communications capabilities of the network equipment in cluster routers 2211a, 2211b, 2211c, the latency and throughput of local cluster networks 2212a, 2212b, 2212c, the latency, throughput, and cost of wide area network links 2213a, 2213b, 2213c, and / or other factors that can contribute to the cost, speed, fault-tolerance, resiliency, efficiency and / or other design criteria of the moderation system architecture.Example Methods of Operation[000228] FIG. 23 is a flowchart of a method 2300, in accordance with example embodiments. Method 2300 can be executed by a computing device, such as computing device 2100. Method 2300 can begin at block 2310, where the method involves identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents.[000229] At block 2320, the method involves detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of corners indicative of a boundary of each candidate document.Atty. Docket: 25-0667-WO[000230] At block 2330, the method involves selecting, by the machine learning model and based on the respective set of comers, one or more documents, of the plurality of candidate documents, that are to be digitally scanned.[000231] At block 2340, the method involves scanning, based on the respective set of corners, the one or more selected documents.[000232] At block 2350, the method involves providing, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.[000233] In some embodiments, the providing of the respective digital version of the scanned one or more documents involves providing a scaled down digital version as each document is scanned.[000234] In some embodiments, the selecting of the one or more documents involves determining a mode of each candidate document without user interaction.[000235] In some embodiments, the detecting of the respective set of corners involves determining a respective confidence score for each respective set of corners, wherein a higher confidence score is indicative of a higher confidence in detecting document corners. The selecting of the one or more documents may be based on the respective confidence scores.[000236] In some embodiments, the scanning of the one or more selected documents involves applying one or more image enhancements to each scanned document. In some embodiments, the one or more image enhancements includes adjusting a contrast, deblurring, dewarping, removal of a finger mark, removal of a shadow, removal of a staple mark, removal of a dog ear, or a combination thereof.[000237] In some embodiments, the machine learning model includes a neural network.[000238] In some embodiments, the neural network includes a MobileNet architecture. Such embodiments involve pruning one or more of the Universal Inverted Blocks to reduce model size.[000239] In some embodiments, the neural network includes a MobileNet architecture with a multi-head attention module.[000240] In some embodiments, the neural network includes a combination of a MobileNet architecture, a Feature Pyramid Network (FPN) decoder, and a CenterNet head.[000241] In some embodiments, the machine learning model includes a bounding box detector configured to detect a document boundary and a corner detector configured to detect document comers.[000242] Some embodiments involve training the machine learning model.Atty. Docket: 25-0667-WO[000243] In some embodiments, the machine learning model may be a student model and the training involves applying knowledge distillation to transfer learning from a teacher model to the machine learning model.[000244] In some embodiments, the training of the machine learning model may be based on one or more of a classification loss, a corner point regression loss, a heatmap loss, or a keypoint offset prediction loss.[000245] In some embodiments, each of the losses may be associated with a respective weight. Such embodiments involve tuning the respective weights to detect documents in one or more of a white-on-white scenario, a document on a document scenario, and a partial occlusion scenario.[000246] In some embodiments, the training of the machine learning model may be based on a synthetic training dataset.[000247] In some embodiments, the synthetic training dataset may have been generated prior to the training of the machine learning model.[000248] In some embodiments, the synthetic training dataset may have been generated by using alpha blending to swap a background of an image of a captured document.[000249] In some embodiments, the synthetic training dataset may have been generated by mixing images of two or more captured documents.[000250] Some embodiments involve determining whether a first document of the selected one or more documents is a duplicate of a second document of the scanned one or more documents, the first document having been identified subsequent to the second document. Such embodiments involve, upon a determination that the first document is a duplicate of the second document, providing, by the display screen of the image capturing device, a user-selectable option to scan the first document.[000251] In some embodiments, the determining whether the first document is a duplicate of the second document involves determining one or more hierarchical feature-level differences between the first document and the second document. Such embodiments involve fusing the one or more hierarchical feature-level differences to combine information across feature-levels. Such embodiments also involve summarizing the combined information across feature-levels. Such embodiments further involve, based on the summarized information, applying a classifier to determine whether the first document is a duplicate of the second document.[000252] In some embodiments, the one or more hierarchical feature-level differences may include differences between low-level textures or edges that appear in the first document and the second document.Atty. Docket: 25-0667-WO[000253] In some embodiments, the one or more hierarchical feature-level differences may include differences between high-level object parts or shapes that appear in the first document and the second document.[000254] In some embodiments, the scanning of the one or more selected documents involves, for each of the plurality of candidate documents, tracking the respective set of corners in consecutive frames of the one or more image frames. Such embodiments involve determining whether a frame stability parameter exceeds a threshold stability value. Such embodiments also involve, upon a determination that the frame stability parameter exceeds the threshold stability value, triggering the scanning of the selected document.[000255] Some embodiments involve determining the frame stability parameter by determining magnitudes of each motion vector corresponding to the tracked respective set of corners. Such embodiments involve determining an aggregate motion metric for the magnitudes of each motion vector. The frame stability parameter is based on the aggregate motion metric.[000256] FIG. 24 is another flowchart of a method 2400, in accordance with example embodiments. Method 2400 can be executed by a computing device, such as computing device 2100. Method 2400 can begin at block 2410, where the method involves detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest.[000257] At block 2420, the method involves tracking the one or more features in consecutive frames of the one or more image frames.[000258] At block 2430, the method involves determining magnitudes of each motion vector corresponding to the tracked one or more features.[000259] At block 2440, the method involves determining an aggregate motion metric for the magnitudes of each motion vector.[000260] At block 2450, the method involves determining a frame stability parameter based on the aggregate motion metric.[000261] At block 2460, the method involves triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.[000262] In some embodiments, the object of interest may be a candidate document to be scanned. The triggering of the capture of the image frame involves scanning the candidate document.[000263] In some embodiments, the one or more features includes a set of comers for the candidate document.Atty. Docket: 25-0667-WO[000264] In some embodiments, the detecting of the one or more features involves determining a respective confidence score for the set of corners, wherein a higher confidence score is indicative of a higher confidence in detecting document corners.[000265] In some embodiments, the detecting of the one or more features may be performed by a machine learning model.[000266] In some embodiments, the machine learning model includes a neural network.[000267] In some embodiments, the neural network includes a MobileNet architecture. Such embodiments involve pruning one or more of the Universal Inverted Blocks to reduce model size.[000268] In some embodiments, the neural network includes a MobileNet architecture with a multi-head attention module.[000269] In some embodiments, the neural network includes a combination of a MobileNet architecture, a Feature Pyramid Network (FPN) decoder, and a CenterNet head.[000270] In some embodiments, the machine learning model includes a bounding box detector configured to detect a document boundary and a corner detector configured to detect document comers.[000271] Some embodiments involve training the machine learning model.[000272] In some embodiments, the machine learning model may be a student model and the training involves applying knowledge distillation to transfer learning from a teacher model to the machine learning model.[000273] In some embodiments, the training of the machine learning model may be based on one or more of a classification loss, a corner point regression loss, a heatmap loss, or a keypoint offset prediction loss.[000274] In some embodiments, each of the losses may be associated with a respective weight. Such embodiments involve tuning the respective weights to detect documents in one or more of a white-on-white scenario, a document on a document scenario, and a partial occlusion scenario.[000275] In some embodiments, the training of the machine learning model may be based on a synthetic training dataset.[000276] In some embodiments, the synthetic training dataset may have been generated prior to the training of the machine learning model.[000277] In some embodiments, the synthetic training dataset may have been generated by using alpha blending to swap a background of an image of a captured document.Atty. Docket: 25-0667-WO[000278] In some embodiments, the synthetic training dataset may have been generated by mixing images of two or more captured documents.[000279] Some embodiments involve determining whether a first object of interest is a duplicate of a second captured object of interest, the first object of interest having been detected subsequent to the second object of interest. Such embodiments involve, upon a determination that the first obj ect of interest is a duplicate of the second captured obj ect of interest, providing, by the display screen of the image capturing device, a user-selectable option to capture the first object of interest.[000280] In some embodiments, the determining whether the first captured object of interest is a duplicate of a second captured object of interest involves determining one or more hierarchical feature-level differences between the first captured object of interest and the second captured object of interest. Such embodiments involve fusing the one or more hierarchical feature-level differences to combine information across feature-levels. Such embodiments also involve summarizing the combined information across feature-levels. Such embodiments further involve, based on the summarized information, applying a classifier to determine whether the first captured object of interest is a duplicate of a second captured object of interest.[000281] In some embodiments, the one or more hierarchical feature-level differences may include differences between low-level textures or edges that appear in the first captured object of interest and the second captured object of interest.[000282] In some embodiments, the one or more hierarchical feature-level differences may include differences between high-level object parts or shapes that appear in the first captured object of interest and the second captured object of interest.[000283] Some embodiments involve determining whether the frame stability parameter exceeds a threshold stability value. The triggering of the capture of the image frame is based upon a determination that the frame stability parameter exceeds the threshold stability value.[000284] The present disclosure is not to be limited in terms of the particular embodiments described in this application, which are intended as illustrations of various aspects. Many modifications and variations can be made without departing from its spirit and scope, as will be apparent to those skilled in the art. Functionally equivalent methods and apparatuses within the scope of the disclosure, in addition to those enumerated herein, will be apparent to those skilled in the art from the foregoing descriptions. Such modifications and variations are intended to fall within the scope of the appended claims.Atty. Docket: 25-0667-WO[000285] The above detailed description describes various features and functions of the disclosed systems, devices, and methods with reference to the accompanying figures. In the figures, similar symbols typically identify similar components, unless context dictates otherwise. The illustrative embodiments described in the detailed description, figures, and claims are not meant to be limiting. Other embodiments can be utilized, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the figures, can be arranged, substituted, combined, separated, and designed in a wide variety of different configurations, all of which are explicitly contemplated herein.[000286] With respect to any or all of the ladder diagrams, scenarios, and flow charts in the figures and as discussed herein, each block and / or communication may represent a processing of information and / or a transmission of information in accordance with example embodiments. Alternative embodiments are included within the scope of these example embodiments. In these alternative embodiments, for example, functions described as blocks, transmissions, communications, requests, responses, and / or messages may be executed out of order from that shown or discussed, including substantially concurrent or in reverse order, depending on the functionality involved. Further, more or fewer blocks and / or functions may be used with any of the ladder diagrams, scenarios, and flow charts discussed herein, and these ladder diagrams, scenarios, and flow charts may be combined with one another, in part or in whole.[000287] A block that represents a processing of information may correspond to circuitry that can be configured to perform the specific logical functions of a herein-described method or technique. Alternatively or additionally, a block that represents a processing of information may correspond to a module, a segment, or a portion of program code (including related data). The program code may include one or more instructions executable by a processor for implementing specific logical functions or actions in the method or technique. The program code and / or related data may be stored on any type of computer readable medium such as a storage device including a disk or hard drive or other storage medium.[000288] The computer readable medium may also include non-transitory computer readable media such as non-transitory computer-readable media that stores data for short periods of time like register memory, processor cache, and random access memory (RAM). The computer readable media may also include non-transitory computer readable media that stores program code and / or data for longer periods of time, such as secondary or persistent longAtty. Docket: 25-0667-WOterm storage, like read only memory (ROM), optical or magnetic disks, compact-disc read only memory (CD-ROM), for example. The computer readable media may also be any other volatile or non-volatile storage systems. A computer readable medium may be considered a computer readable storage medium, for example, or a tangible storage device.[000289] Moreover, a block that represents one or more information transmissions may correspond to information transmissions between software and / or hardware modules in the same physical device. However, other information transmissions may be between software modules and / or hardware modules in different physical devices.[000290] While various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are provided for explanatory purposes and are not intended to be limiting, with the true scope being indicated by the following claims.

Claims

Atty. Docket: 25-0667-WOCLAIMSWhat is claimed is:

1. A computer-implemented method, comprising:identifying, in one or more image frames of a live preview of an image capturing device, a presence of a plurality of candidate documents;detecting, by a machine learning model and for each of the plurality of candidate documents, a respective set of comers indicative of a boundary of each candidate document;selecting, by the machine learning model and based on the respective set of corners, one or more documents, of the plurality of candidate documents, that are to be digitally scanned;scanning, based on the respective set of corners, the one or more selected documents; andproviding, by a display screen of the image capturing device, a respective digital version of the scanned one or more documents.

2. The computer-implemented method of claim 1, wherein the providing of the respective digital version of the scanned one or more documents comprises providing a scaled down digital version as each document is scanned.

3. The computer-implemented method of claim 1 , wherein the selecting of the one or more documents comprises determining a mode of each candidate document without user interaction.

4. The computer-implemented method of claim 1, wherein the detecting of the respective set of corners further comprises:determining a respective confidence score for each respective set of corners, wherein a higher confidence score is indicative of a higher confidence in detecting document corners, and wherein the selecting of the one or more documents is based on the respective confidence scores.

5. The computer-implemented method of claim 1 , wherein the scanning of the one or more selected documents further comprises:Atty. Docket: 25-0667-WOapplying one or more image enhancements to each scanned document.

6. The computer-implemented method of claim 5, wherein the one or more image enhancements comprises adjusting a contrast, deblurring, dewarping, removal of a finger mark, removal of a shadow, removal of a staple mark, removal of a dog ear, or a combination thereof.

7. The computer-implemented method of claim 1, wherein the machine learning model comprises a neural network.

8. The computer-implemented method of claim 7, wherein the neural network comprises a MobileNet architecture, and further comprising:pruning one or more of the Universal Inverted Blocks to reduce model size.

9. The computer-implemented method of claim 7, wherein the neural network comprises a MobileNet architecture with a multi-head attention module.

10. The computer-implemented method of claim 7, wherein the neural network comprises a combination of a MobileNet architecture, a Feature Pyramid Network (FPN) decoder, and a CenterNet head.

11. The computer-implemented method of claim 1, wherein the machine learning model comprises a bounding box detector configured to detect a document boundary and a comer detector configured to detect document comers.

12. The computer-implemented method of claim 1, further comprising:training the machine learning model.

13. The computer-implemented method of claim 12, wherein the machine learning model is a student model and the training comprises applying knowledge distillation to transfer learning from a teacher model to the machine learning model.

14. The computer-implemented method of claim 12, wherein the training of the machine learning model is based on one or more of a classification loss, a corner point regression loss, a heatmap loss, or a keypoint offset prediction loss.Atty. Docket: 25-0667-WO15. The computer-implemented method of claim 14, wherein each of the losses is associated with a respective weight, and further comprising:tuning the respective weights to detect documents in one or more of a white-on-white scenario, a document on a document scenario, and a partial occlusion scenario.

16. The computer-implemented method of claim 12, wherein the training of the machine learning model is based on a synthetic training dataset.

17. The computer-implemented method of claim 16, the synthetic training dataset having been generated prior to the training of the machine learning model.

18. The computer-implemented method of claim 17, the synthetic training dataset having been generated by using alpha blending to swap a background of an image of a captured document.

19. The computer-implemented method of claim 17, the synthetic training dataset having been generated by mixing images of two or more captured documents.

20. The computer-implemented method of claim 1, further comprising:determining whether a first document of the selected one or more documents is a duplicate of a second document of the scanned one or more documents, the first document having been identified subsequent to the second document; andupon a determination that the first document is a duplicate of the second document, providing, by the display screen of the image capturing device, a user-selectable option to scan the first document.

21. The computer-implemented method of claim 20, wherein the determining whether the first document is a duplicate of the second document further comprises:determining one or more hierarchical feature-level differences between the first document and the second document;fusing the one or more hierarchical feature-level differences to combine information across feature-levels;summarizing the combined information across feature-levels; andAtty. Docket: 25-0667-WObased on the summarized information, applying a classifier to determine whether the first document is a duplicate of the second document.

22. The computer-implemented method of claim 21, wherein the one or more hierarchical feature-level differences comprise differences between low-level textures or edges that appear in the first document and the second document.

23. The computer-implemented method of claim 21, wherein the one or more hierarchical feature-level differences comprise differences between high-level object parts or shapes that appear in the first document and the second document.

24. The computer-implemented method of claim 1 , wherein the scanning of the one or more selected documents further comprises:for each of the plurality of candidate documents, tracking the respective set of comers in consecutive frames of the one or more image frames;determining whether a frame stability parameter exceeds a threshold stability value; andupon a determination that the frame stability parameter exceeds the threshold stability value, triggering the scanning of the selected document.

25. The computer-implemented method of claim 24, further comprising:determining the frame stability parameter by:determining magnitudes of each motion vector corresponding to the tracked respective set of corners;determining an aggregate motion metric for the magnitudes of each motion vector, andwherein the frame stability parameter is based on the aggregate motion metric.

26. A computer-implemented method, comprising:detecting, in one or more image frames of a live preview of an image capturing device, one or more features corresponding to an object of interest;tracking the one or more features in consecutive frames of the one or more image frames;Atty. Docket: 25-0667-WOdetermining magnitudes of each motion vector corresponding to the tracked one or more features;determining an aggregate motion metric for the magnitudes of each motion vector; determining a frame stability parameter based on the aggregate motion metric; and triggering, based on the frame stability parameter, a capture of an image frame comprising the object of interest.

27. The computer-implemented method of claim 26, wherein the object of interest is a candidate document to be scanned, and wherein the triggering of the capture of the image frame comprises scanning the candidate document.

28. The computer-implemented method of claim 27, wherein the one or more features comprises a set of corners for the candidate document.

29. The computer-implemented method of claim 28, wherein the detecting of the one or more features further comprises:determining a respective confidence score for the set of corners, wherein a higher confidence score is indicative of a higher confidence in detecting document corners.

30. The computer-implemented method of claim 28, wherein the detecting of the one or more features is performed by a machine learning model.

31. The computer-implemented method of claim 30, wherein the machine learning model comprises a neural network.

32. The computer-implemented method of claim 31, wherein the neural network comprises a MobileNet architecture, and further comprising:pruning one or more of the Universal Inverted Blocks to reduce model size.

33. The computer-implemented method of claim 31, wherein the neural network comprises a MobileNet architecture with a multi-head attention module.Atty. Docket: 25-0667-WO34. The computer-implemented method of claim 31, wherein the neural network comprises a combination of a MobileNet architecture, a Feature Pyramid Network (FPN) decoder, and a CenterNet head.

35. The computer-implemented method of claim 30, wherein the machine learning model comprises a bounding box detector configured to detect a document boundary and a comer detector configured to detect document comers.

36. The computer-implemented method of claim 30, further comprising:training the machine learning model.

37. The computer-implemented method of claim 36, wherein the machine learning model is a student model and the training comprises applying knowledge distillation to transfer learning from a teacher model to the machine learning model.

38. The computer-implemented method of claim 36, wherein the training of the machine learning model is based on one or more of a classification loss, a corner point regression loss, a heatmap loss, or a keypoint offset prediction loss.

39. The computer-implemented method of claim 38, wherein each of the losses is associated with a respective weight, and further comprising:tuning the respective weights to detect documents in one or more of a white-on-white scenario, a document on a document scenario, and a partial occlusion scenario.

40. The computer-implemented method of claim 36, wherein the training of the machine learning model is based on a synthetic training dataset.

41. The computer-implemented method of claim 40, the synthetic training dataset having been generated prior to the training of the machine learning model.

42. The computer-implemented method of claim 41, the synthetic training dataset having been generated by using alpha blending to swap a background of an image of a captured document.Atty. Docket: 25-0667-WO43. The computer-implemented method of claim 41, the synthetic training dataset having been generated by mixing images of two or more captured documents.

44. The computer-implemented method of claim 26, further comprising:determining whether a first object of interest is a duplicate of a second captured object of interest, the first object of interest having been detected subsequent to the second object of interest; andupon a determination that the first object of interest is a duplicate of the second captured object of interest, providing, by the display screen of the image capturing device, a user-selectable option to capture the first object of interest.

45. The computer-implemented method of claim 44, wherein the determining whether the first captured object of interest is a duplicate of a second captured object of interest further comprises:determining one or more hierarchical feature-level differences between the first captured object of interest and the second captured object of interest;fusing the one or more hierarchical feature-level differences to combine information across feature-levels;summarizing the combined information across feature-levels; andbased on the summarized information, applying a classifier to determine whether the first captured object of interest is a duplicate of a second captured object of interest.

46. The computer-implemented method of claim 45, wherein the one or more hierarchical feature-level differences comprise differences between low-level textures or edges that appear in the first captured object of interest and the second captured object of interest.

47. The computer-implemented method of claim 45, wherein the one or more hierarchical feature-level differences comprise differences between high-level object parts or shapes that appear in the first captured object of interest and the second captured object of interest.

48. The computer-implemented method of claim 26, further comprising:determining whether the frame stability parameter exceeds a threshold stability value, andAtty. Docket: 25-0667-WOwherein the triggering of the capture of the image frame is based upon a determination that the frame stability parameter exceeds the threshold stability value.

49. A computing device, comprising:one or more processors; anddata storage, wherein the data storage has stored thereon computer-executable instructions that, when executed by the one or more processors, cause the computing device to carry out functions comprising the computer-implemented method of any one of claims 1-48.

50. A computer program comprising instructions that, when executed by a computer, cause the computer to perform steps in accordance with the method of any one of claims 1-48.

51. An article of manufacture comprising one or more non-transitory computer readable media having computer-readable instructions stored thereon that, when executed by one or more processors of a computing device, cause the computing device to carry out functions that comprise the computer-implemented method of any one of claims 1-48.

52. A system comprising instructions that, when executed by a computer, cause the computer to perform steps in accordance with the method of any one of claims 1-48.