Self-adaptive perceptual archive digital scanning processing system
Through the adaptive sensing archive digital scanning processing system, using deep learning and multimodal sensor technology, the problem that traditional scanning methods cannot adapt to complex archive content is solved, and high-quality and efficient archive digital processing is achieved, suitable for a variety of document types and conditions.
Patent Information
- Application Number
- CN202411823611.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-01-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional digital scanning methods for archives lack flexibility and cannot adapt to complex and diverse archive content, resulting in uneven scanning quality and inefficient efficiency, especially when processing archives containing complex charts, pictures or handwritten text, which may lead to blurred images or loss of information.
Adaptive sensing archive digital scanning processing system, combined with deep learning dynamic sensing document content complexity, through deep reinforcement learning model and multimodal sensor fusion technology, the best scanning mode and data fusion strategy are automatically selected to achieve document refinement Scan and automated processing.
It improves the quality and efficiency of digital scans of archives, ensures the optimization of processing processes under different document types and conditions, ensures the accuracy of image quality and information extraction, and supports multiple format output and metadata embedding.
Smart Images
Figure CN119316525A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence applications, and in particular to an adaptive perception archive digital scanning and processing system. Background Art
[0002] With the explosive growth of information in modern society, the number and complexity of archives have not only increased dramatically in quantity, but also become increasingly diversified in content and form, which has brought unprecedented challenges to the digital management of archives.
[0003] Although the commonly used scanning methods have been the main means of archive digitization for some time in the past, their inherent fixedness and lack of flexibility make it difficult to adapt to the current complex and changing archive processing needs. Specifically, traditional scanning methods usually rely on preset scanning parameters, such as fixed resolution, color mode, etc., which are often unable to be flexibly adjusted according to the actual situation of the archives. Therefore, when processing a large number of archives, traditional scanning methods often have problems such as uneven scanning quality and low efficiency.
[0004] In addition, with the diversification of archive content, traditional scanning methods are also facing new challenges. For example, for archives containing complex charts, pictures or handwritten text, traditional scanning methods may not be able to accurately capture these details, resulting in blurred scanned images or information loss. This not only affects the digitization quality of archives, but also brings many inconveniences to the subsequent use and management of archives. Summary of the invention
[0005] In view of the deficiencies in the prior art, the present invention provides an adaptive perception archive digital scanning and processing system. The system aims to dynamically perceive the complexity of the document content through deep learning, automatically select the most appropriate scanning mode and data fusion strategy, and thus improve the quality and efficiency of archive digital scanning. By using deep reinforcement learning (DRL), multimodal sensor fusion technology, and advanced image processing algorithms, refined scanning and automated processing of documents are achieved, ensuring that the system can adaptively optimize the processing process under different document types and conditions.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions: An adaptive sensing archive digital scanning and processing system, comprising: An input module, which is used to receive archival documents to be scanned, supports manual placement and / or automatic conveyor belt input, and can adapt to various document types; The pre-processing module performs preliminary processing on the received documents, including dust removal, wrinkle removal and alignment, to improve the accuracy of subsequent scanning; The perception module uses a deep reinforcement learning model and a sensor array to achieve dynamic perception and adaptive scanning, and selects the best scanning method in real time based on the complexity, quality, and type of the document content; Data fusion module, which fuses data captured by multiple sensors to optimize image quality and information extraction; The post-processing module further processes the scanned images to improve image quality and information extraction accuracy; The output module outputs the processed digital archives to the specified storage location, supporting multiple formats and metadata embedding.
[0007] Preferably, the input module further comprises: Automatic document feeder to support automatic delivery of large batches of documents, reducing manual intervention; and Physical sensors, combined with barcode scanning and QR code recognition functions, automatically identify the classification information of files.
[0008] Preferably, the deep reinforcement learning model in the perception module includes a state space S, an action space A and a reward function R(s,a), wherein the state space S describes various features of the document, the action space A includes scanning operations that the system can take, and the reward function R(s,a) is determined based on the quality of the scanning results.
[0009] Preferably, the specific calculation process of the reward function is as follows: the quality of the scanned image can be measured by the following factors: clarity C, noise level N, image integrity I, then the reward function is: R(s,a)=α·C(s,a)-β·N(s,a)+γ·I(s,a) in: C(s,a): The clarity of the scanned image, usually calculated by image sharpness or contrast, the higher the better; N(s,a): Image noise level, calculated by the image noise index PSNR, the lower the better; I(s,a): Image integrity, indicating whether the scanned area of the document is complete, the higher the better; α, β, γ: are the weight coefficients of clarity, noise level and completeness respectively.
[0010] Preferably, the deep reinforcement learning model updates the Q value in the following manner:
[0011] in: Q(s t ,a t ) means in state a t Next, perform action a tQ value; α represents the learning rate, which controls the speed of Q value update; r t represents the reward received at time step t; γ represents the discount factor, controlling the importance of future rewards; Indicates the next state s t+1 The maximum Q value under , represents the expected maximum reward.
[0012] Preferably, the sensor array includes an optical scanning sensor, an infrared scanning sensor and a laser scanning sensor, which are respectively used to capture basic image information, hidden information and 3D depth information of the document.
[0013] Preferably, the data fusion module uses weighted average fusion, principal component analysis or maximum and minimum selection method to fuse data from different sensors to optimize image quality and information extraction.
[0014] Preferably, when principal component analysis is used for data fusion, the following steps are included: a. Centralize the sensor data: Subtract the mean from each sensor data dimension to obtain the centralized data matrix X'. The formula is as follows: X'=X-μ Where the sensor data matrix is X, μ is the mean vector of each column; b. Calculate the covariance matrix of the centralized data:
[0015] Where ∑ is the covariance matrix, n is the number of samples, X ,T is the transpose of the centered data matrix; c. Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues: d. Select the first few principal components with larger eigenvalues to form a new low-dimensional space. Finally, the fused data can be represented by the principal components after dimensionality reduction: X PCA =X'·W Among them, W is the selected principal component matrix.
[0016] Preferably, the post-processing module further includes denoising and image enhancement technical functions, OCR recognition technical functions and region segmentation and labeling functions, wherein denoising and image enhancement are performed using convolutional neural networks, OCR recognition is used to extract text information in documents, and region segmentation and labeling are used to identify and label important areas in complex documents.
[0017] Preferably, the output module supports outputting archives to common formats such as PDF, TIFF, JPEG, etc., and can embed metadata as needed, while supporting file management on cloud storage platforms as well as local storage.
[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides an adaptive perception archive digital scanning and processing system, which combines advanced technologies such as deep reinforcement learning, sensor array, and data fusion to achieve intelligent and automated archive digital scanning and processing. The system can automatically select the best scanning method according to the complexity, quality, and type of the document content, improve the scanning quality and processing efficiency, and ensure that the digital content of the archive is completely and accurately restored. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0020] Figure 1 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION
[0021] The present invention is further described in detail below with reference to the accompanying drawings.
[0022] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be used for other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.
[0023] See also Figure 1 The present invention provides an adaptive perception archive digital scanning and processing system, comprising: An input module, which is used to receive archival documents to be scanned, supports manual placement and / or automatic conveyor belt input, and can adapt to various document types; The pre-processing module performs preliminary processing on the received documents, including dust removal, wrinkle removal and alignment, to improve the accuracy of subsequent scanning; The perception module uses a deep reinforcement learning model and a sensor array to achieve dynamic perception and adaptive scanning, and selects the best scanning method in real time based on the complexity, quality, and type of the document content; Data fusion module, which fuses data captured by multiple sensors to optimize image quality and information extraction; The post-processing module further processes the scanned images to improve image quality and information extraction accuracy; The output module outputs the processed digital archives to the specified storage location, supporting multiple formats and metadata embedding.
[0024] The following is a detailed description of the functions of each module of this system: The input module of the present invention includes an automatic document feeder (ADF) and a physical sensor; it is responsible for receiving the archive documents to be scanned and supports input methods such as manual placement and automatic conveyor belt. It can adapt to various document types, including paper documents, photos, drawings, handwritten materials, etc. Specifically:
[0025] Automatic Document Feeder (ADF): supports automatic transmission of large batches of documents, reducing manual intervention; Physical sensor: Combines barcode scanning, QR code recognition and other functions to automatically identify the classification information of files.
[0026] The pre-processing module of the present invention has the functions of image repair and document alignment, and is used to perform preliminary processing on the received documents, including dust removal, wrinkle removal and alignment, so as to improve the accuracy of subsequent scanning. Specifically:
[0027] Image restoration: Use image processing algorithms to remove dust and wrinkles from the document surface, reducing the impact of environmental factors (such as paper creases) on the scanning effect; Document Alignment: Uses image geometry correction technology to ensure that every part of the scanned document is aligned.
[0028] The perception module of the present invention is the core of the entire system. It realizes dynamic perception and adaptive scanning through deep reinforcement learning model and sensor array, and selects the best scanning mode in real time according to the complexity, quality and type of document content. Specifically:
[0029] (1) The deep reinforcement learning model includes a data collection and annotation unit, a model training unit, and a real-time decision-making unit. The deep reinforcement learning model selects the best scanning method in real time based on the complexity, quality, and type of the document content.
[0030] Data Collection and Annotation Unit: Collect document samples of varying complexity (such as images, text, graffiti, damaged parts, etc.) and manually annotate them. The annotation content includes document type, regional features (such as font, color, image type, etc.) and optimal scanning method.
[0031] Model training unit: Use deep reinforcement learning algorithms (such as Proximal Policy Optimization, PPO or Deep Q-Network, DQN) to train the model so that it can learn how to select the appropriate scanning method based on the characteristics of the document.
[0032] Real-time decision unit: Based on the real-time feedback of the document (e.g., image quality, regional features, etc.), the model decides whether to switch the scanning mode (e.g., select optical scanning, infrared scanning, laser scanning, etc.) and optimize the scanning parameters (e.g., resolution, exposure time, color adjustment, etc.).
[0033] Among them, deep reinforcement learning models include 1. Basic framework of deep reinforcement learning (DRL) Deep Reinforcement Learning (DRL) combines deep learning and reinforcement learning, enabling intelligent systems to learn the best decision-making strategy from the environment. In this solution, the DRL model is used to dynamically select the appropriate scanning method and adjust the scanning parameters.
[0034] State space S: describes various characteristics of the document, such as document type (text, image, graffiti), regional characteristics (complexity, degree of damage, etc.), and environmental factors (light, temperature and humidity, etc.).
[0035] Action space A: The scanning operations that the system can take, such as selecting scanning resolution, exposure time, scanning mode (optical, infrared, laser), etc.
[0036] Reward function R(s,a): The reward obtained after performing an action. The reward function is determined based on the quality of the scan results (e.g. image clarity, noise level).
[0037] 2. Calculation of reward function In deep reinforcement learning, the goal of the model is to maximize the cumulative reward. In order to select the best scanning pattern and parameters, a suitable reward function needs to be designed.
[0038] The quality of a scanned image can be measured by the following factors: Clarity C, noise level N, image integrity (whether the document area is complete) I; The reward function can be defined as: R(s,a)=α·C(s,a)-β·N(s,a)+γ·I(s,a) in: C(s,a): The clarity of the scanned image, usually calculated by image sharpness or contrast, the higher the better; N(s,a): Image noise level, calculated by the image noise index (such as PSNR, PeakSignaltoNoiseRatio), the lower the better; I(s,a): Image integrity, indicating whether the scanned area of the document is complete and not blocked or damaged. The higher the better. α, β, γ: are the weight coefficients of clarity, noise and completeness respectively.
[0039] Through training, the deep reinforcement learning model can learn to automatically select appropriate scanning parameters based on the state of the document to maximize the reward.
[0040] 3. Deep Q-Network (DQN) Update Rules The Q-learning algorithm is used to optimize the policy, and the Q function is used to represent the value of the state-action pair, that is, the expected cumulative reward for performing an action in a given state.
[0041] The formula for updating the Q value is:
[0042] in: Q(s t ,a t ): In state s t Next, perform action a t Q value; α: learning rate, which controls the speed of Q value update; r t : The reward received at time step t; γ: Discount factor, controlling the importance of future rewards.
[0043] : Next state s t+1 The maximum Q value under , represents the expected maximum reward. The goal of this algorithm is to learn the optimal strategy by continuously adjusting the Q value, that is, to maximize the effect of document scanning.
[0044] (2) The sensor array consists of optical scanning sensors, infrared scanning sensors, laser scanning sensors and other devices; it is used to capture multimodal information of the document, including conventional image information, infrared information, 3D depth information, etc.
[0045] Optical scanning sensor: A high-resolution optical sensor is used to capture basic image information of the document; Infrared scanning sensors: used to identify hidden graffiti, alterations or watermarks, especially for historical archives and document recovery; Laser scanning sensor: used to obtain 3D depth information of documents, suitable for scanning complex charts, cultural relics, ancient books or damaged documents.
[0046] The data fusion module of the present invention adopts weighted average fusion, principal component analysis (PCA) or maximum and minimum value selection method; it is used to fuse the data captured by multiple sensors to optimize image quality and information extraction.
[0047] Weighted average fusion: Based on the accuracy of the sensor and the specific requirements of the document, a weighted average strategy is used to merge the data of different sensors to ensure that the advantages of each sensor are maximized; Principal Component Analysis (PCA): PCA is used to reduce the dimension of sensor data, remove redundant information, retain key information in the document, and improve data fusion effect; Max-min selection: In some cases, the best sensor output (such as depth information captured using laser scanning) is selected to further improve image accuracy.
[0048] Furthermore, the data fusion module combines information from different types of sensors (optical, infrared, laser, etc.) to improve the scanning accuracy and information integrity. In this solution, the principal component analysis (PCA) data fusion strategy is adopted:
[0049] Principal component analysis (PCA) is a dimensionality reduction technique that can map multidimensional sensor data to a lower dimensional space, thereby retaining the main features. The goal of PCA is to diagonalize the covariance matrix of sensor data through orthogonal transformation and extract the principal components that best characterize the data variation.
[0050] Let the sensor data matrix be X, where each row represents a sample and each column represents a different feature. The PCA calculation process is as follows:
[0051] 1. Data centering: Subtract the mean from each sensor data dimension to obtain the centralized data matrix X'; X'=X-μ Here, μ is the mean vector of each column.
[0052] 2. Calculate the covariance matrix:
[0053] Where ∑ is the covariance matrix, n is the number of samples, X ,T is the transpose of the centered data matrix.
[0054] 3. Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues. The eigenvectors correspond to the main components of the data, and the eigenvalues indicate the importance of each principal component.
[0055] 4. Select principal components: Select the first few principal components with larger eigenvalues to form a new low-dimensional space. Finally, the fused data can be represented by the principal components after dimensionality reduction:
[0056] X PCA =X'·W Among them, W is the selected principal component matrix.
[0057] The post-processing module of the present invention includes denoising and image enhancement functions, OCR recognition functions, and region segmentation and annotation functions; and is used to further process the scanned image to improve image quality and information extraction accuracy.
[0058] Denoising and image enhancement: Use convolutional neural network (CNN) for denoising and image enhancement to improve image clarity and readability; OCR recognition: extract text information from documents through OCR (optical character recognition) technology and perform semantic understanding; Region segmentation and annotation technology: In complex documents, image segmentation technology is used to identify and annotate important areas (such as signatures, diagrams, and text) to ensure the accuracy of subsequent processing and retrieval.
[0059] The output module of the present invention includes multi-format output function, cloud storage and local storage support function; it is used to output the processed digital archives to a designated storage location, and supports multiple formats and metadata embedding.
[0060] Multi-format output: supports outputting archives to common formats such as PDF, TIFF, JPEG, etc., and can embed metadata (such as document source, creation time, classification information, etc.) as needed; Cloud storage and local storage support: Supports file management on cloud storage platforms and local storage to facilitate document management and sharing.
[0061] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.
Claims
1. An adaptive perception archive digital scanning and processing system, characterized by: include, An input module, which is used to receive archival documents to be scanned, supports manual placement and / or automatic conveyor input, and can adapt to various document types; The pre-processing module performs preliminary processing on the received documents, including dust removal, wrinkle removal and alignment, to improve the accuracy of subsequent scanning; The perception module uses a deep reinforcement learning model and a sensor array to achieve dynamic perception and adaptive scanning, and selects the best scanning method in real time based on the complexity, quality, and type of the document content; Data fusion module, which fuses data captured by multiple sensors to optimize image quality and information extraction; The post-processing module further processes the scanned images to improve image quality and information extraction accuracy; Output module, which outputs the processed digital archives to the specified storage location, supporting multiple formats and metadata embedding; Among them, the quality of the scanned image can be measured by the following factors: clarity C, noise level N, image integrity I; The sensor array includes an optical scanning sensor, an infrared scanning sensor and a laser scanning sensor, which are respectively used to capture basic image information, hidden information and 3D depth information of the document; Among them, the post-processing module further processes functions including denoising and image enhancement technology, OCR recognition technology and region segmentation and annotation functions, among which denoising and image enhancement are performed using convolutional neural networks, OCR recognition is used to extract text information in documents, and region segmentation and annotation is used to identify and annotate important areas in complex documents.
2. The adaptive sensing archive digital scanning and processing system according to claim 1, characterized in that: The input module further comprises: Automatic document feeder to support automatic delivery of large batches of documents, reducing manual intervention; and Physical sensors, combined with barcode scanning and QR code recognition functions, automatically identify the classification information of files.
3. The adaptive sensing archive digital scanning and processing system according to claim 1, characterized in that: The deep reinforcement learning model in the perception module includes a state space S, an action space A and a reward function R(s,a), wherein the state space S describes various features of the document, the action space A includes scanning operations that the system can take, and the reward function R(s,a) is determined according to the quality of the scanning results.
4. The adaptive sensing archive digital scanning and processing system according to claim 3, characterized in that: The specific calculation process of the reward function is as follows: The reward function R(s,a) is: R(s,a)=α·C(s,a)-β·N(s,a)+γ·I(s,a) in: C(s,a): The clarity of the scanned image, usually calculated by image sharpness or contrast, the higher the better; N(s,a): Image noise level, calculated by the image noise index PSNR, the lower the better; I(s,a): Image integrity, indicating whether the scanned area of the document is complete, the higher the better; α, β, γ: are the weight coefficients of clarity, noise level and completeness respectively.
5. The adaptive sensing archive digital scanning and processing system according to claim 4, characterized in that: The deep reinforcement learning model updates the Q value in the following way: ； in: Q(s t ,a t ) means in state a t Next, perform action a t Q value; α represents the learning rate, which controls the speed of Q value update; r t represents the reward received at time step t; γ represents the discount factor, controlling the importance of future rewards; Indicates the next state s t+1 The maximum Q value under , represents the expected maximum reward.
6. The adaptive sensing archive digital scanning and processing system according to claim 1, characterized in that: The data fusion module uses weighted average fusion, principal component analysis or maximum and minimum value selection method to fuse data from different sensors to optimize image quality and information extraction.
7. The adaptive sensing archive digital scanning and processing system according to claim 6, characterized in that: When principal component analysis is used for data fusion, the following steps are included: a. Centralize the sensor data: Subtract the mean from each sensor data dimension to obtain the centralized data matrix X'. The formula is as follows: X'=X-μ Where the sensor data matrix is X, μ is the mean vector of each column; b. Calculate the covariance matrix of the centralized data: ； Where ∑ is the covariance matrix, n is the number of samples, X ,T is the transpose of the centered data matrix; c. Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues; d. Select the first few principal components with larger eigenvalues to form a new low-dimensional space. Finally, the fused data can be represented by the principal components after dimensionality reduction: X PCA =X'·W; Among them, W is the selected principal component matrix.
8. The adaptive sensing archive digital scanning and processing system according to claim 1, characterized in that: The output module supports the output of archives into common formats such as PDF, TIFF, and JPEG, and can embed metadata as needed. It also supports file management on cloud storage platforms and local storage.