Multifunctional standardized acquisition system and device for railway engineering image data

Abstract

The invention relates to the technical field of data processing, and discloses a multifunctional standardized acquisition system and device for railway engineering image data, which realize standardization of railway engineering image data acquisition through parallel processing among multiple units, ensure real-time updating and verification of field information, strengthen identity authentication of personnel and equipment and improve the efficiency of data processing. Human errors and data distortion are avoided, and the overall reliability and safety of the system are improved; and finally, through regular timing, re-checking the text recognized by the OCR, performing high-precision face recognition, and unifying identity verification of equipment and personnel, time synchronization, text accuracy and identity authentication tightness of the system are ensured, the overall precision and security of the system are improved, potential loopholes or identity false use are prevented, and the system is safe and reliable. Normative management of a construction site and authenticity of information are guaranteed; the whole system is intelligent and portable, and accurate analysis and response rate of the system are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of image acquisition technology, and more specifically, to a multi-functional standardized acquisition system and device for railway engineering image data. Background Technology

[0002] Railway engineering imagery data is acquired during railway construction by using various imaging devices (such as drones, aerial photography, and ground-based photography) to capture real-time or periodic images of the construction site, progress, and quality. This imagery data includes, but is not limited to, high-resolution photographs, videos, and satellite images, effectively recording each stage of the project and serving for later analysis, monitoring, evaluation, and record management. Imagery data acquisition is of great significance for quality control, safety management, and progress monitoring in railway projects.

[0003] However, since the implementation of the management documents, existing technologies have faced numerous challenges in the collection of railway engineering image data. For example, unclear personnel descriptions, incomplete acceptance team members in video footage, illegible handwritten text on signs, non-standard naming of image data, and blurry image quality. Among the existing technologies for data collection, there is no system for the standardized collection and management of engineering image data, which makes it difficult to implement the requirements of the management documents. This further restricts the overall planning and long-term development of railway projects.

[0004] Based on this, the present invention proposes a multi-functional standardized acquisition system and device for railway engineering image data to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a multi-functional standardized acquisition system and device for railway engineering image data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a multi-functional standardized acquisition system for railway engineering image data, the system comprising: an image integration unit, a content detection unit, and a data archiving unit connected in sequence, wherein: The image integration unit is used to assist in guiding the acquisition of images and sound sources, perform preliminary face processing and text recognition on the acquired images and output initial field-level confidence and initial face confidence, as well as pre-name and sign the images. The content detection unit is used to receive low-confidence data output by the image integration unit, perform secondary recognition of the sign text in the image, extract facial features, and perform regular expression verification based on the naming. The data archiving unit is used to perform final verification and archiving of low-confidence data processed by the content detection unit, as well as to independently verify identity and unify identity source.

[0007] Furthermore, the image integration unit includes: a shooting reminder module, an auxiliary guidance module, a clear voice reminder module, a text and face preprocessing module, and a pre-naming module, wherein: The shooting reminder module integrates a light sensor, a magnetic field sensor, and an acceleration sensor, and combines them with a direction rotation matrix function to convert sensor data into direction information, thereby detecting whether the image is tilted. The auxiliary guidance module deploys a YOLO-nano-Steel detection model to detect the presence of measuring instruments. The voice clarity reminder module: captures the sound source through a microphone, and uses noise reduction technology and PicoTTS voice engine to perform linear quantization to obtain the voice clarity; The text and face preprocessing module is used to process images, generate cropped images and convert them into grayscale images, then perform text detection and pre-recognition on the grayscale images, and filter faces in the images in real time. The pre-naming module: obtains naming rules, pre-names images, and performs rapid self-checks using local regular expressions and enumeration values.

[0008] Furthermore, the voice clarity reminder module: captures the sound source through a microphone, and uses noise reduction technology and the PicoTTS voice engine to perform linear quantization to obtain voice clarity, specifically as follows: A1: Real-time audio data acquisition via microphone at a sampling rate of 16kHz and in mono mode; A2: Real-time noise reduction using a recurrent neural network (RNN); A3: Further processing is performed using a 1D-CNN convolutional neural network to extract sampling points and perform linear quantization to obtain speech intelligibility; Specifically: During the linear quantization process, a quantization scoring threshold is set, and the speech clarity is compared with the quantization scoring threshold; If the speech clarity is less than 50, it is judged as unclear, and a prompt to repeat is made; If 50 ≤ speech clarity < 70, it is judged as unclear, and a prompt is made to repeat it closer; If 70 ≤ speech clarity, the condition is passed, and silence is maintained.

[0009] Furthermore, the content detection unit includes: a secondary text recognition module, a face feature extraction module, and a naming verification module, wherein: The secondary text recognition module introduces an adaptive scale fusion structure to improve robustness and performs secondary recognition on the pre-recognized text. The facial feature extraction module detects key facial features within the image and extracts feature maps; The naming verification module: After the pre-naming module uploads to the naming verification module, it uses local regular expressions and enumeration values ​​to perform deduplication detection again, generates a compressed file, and uploads it.

[0010] Furthermore, the data archiving unit includes: a naming check module, a content verification module, and a face detection module, wherein: The naming check module receives compressed files, adds a trust flag, and then stores them in the database. The content verification module further re-examines the questionable cropped images, outputs a text sequence, and performs regular expression verification on the dates. The face detection module enhances the fitting ability for irregularly shaped objects, introduces weighting coefficients, suppresses weakly correlated background information, and verifies faces within the image.

[0011] Furthermore, the content verification module further re-examines the questionable cropped images, outputs a text sequence, and performs regular expression validation on the dates. The specific operations are as follows: B1: Obtain the initial field-level confidence scores of each module, and assign weights to each module using a static weighting strategy. The weight of the text-face preprocessing module is 0, the weight of the secondary text recognition module is 0.7, and the weight of the content verification module is 0.3. B2: The field-level confidence score of the text is calculated based on the weights of each module, and the expression is as follows: F = 0.7 × se + 0.3 × sc; In the formula: se is the initial field-level confidence score output by the secondary text recognition module, and sc is the field-level confidence score output by the content verification module; B3: Set the threshold for determining the confidence level of the merged field: F≥0.95 is acceptable; F<0.95 triggers manual intervention.

[0012] Furthermore, the process of obtaining the initial field-level confidence scores for each module is as follows: C1: The initial field-level confidence of the text-face preprocessing module is obtained by evaluating the cropped image using the CRNN-Lite lightweight convolutional recurrent neural network; If the initial field-level confidence score of the text-face preprocessing module is >0.9, it can pass; if the initial field-level confidence score of the text-face preprocessing module is <0.9, it will be uploaded to the secondary text recognition module. C2: The secondary text recognition module re-detects and recognizes the cropped image and outputs the initial field-level confidence score. If the output initial field-level confidence score is ≥0.98, it passes and is not uploaded to the content verification module. If the output initial field-level confidence score is <0.98, it is uploaded to the content verification module. C3: The content verification module performs accuracy checks on questionable cropped images, provides an initial field-level confidence score, and calculates the merged field-level confidence score.

[0013] Furthermore, the face detection module enhances its ability to fit irregularly shaped objects by introducing weighting coefficients, suppressing weakly correlated background information, and verifying faces within the image. The specific operations are as follows: D1: Obtain the initial face confidence scores of each module. The same static weighting strategy is used to assign weights to each module. The weight of the text face preprocessing module is 0, the weight of the face feature extraction module is 0.7, and the weight of the face detection module is 0.3. D2: Calculated based on the initial face confidence scores of each module, expressed as: F = 0.70 × se + 0.30 × sc; In the formula: se is the highest cosine similarity obtained by the face feature extraction module after comparing with the local index; sc represents the highest cosine similarity obtained by the face detection module after comparing with the global face database.

[0014] Furthermore, the process of obtaining the initial face confidence scores of each module is as follows: E1: If the initial confidence level output by the text-face preprocessing module is <0.7, it is discarded and the upload is abandoned. If the initial confidence level output by the text-face preprocessing module is ≥0.7, it is uploaded to the face feature extraction module. E2: The face feature extraction module detects the extracted feature map to obtain an initial face confidence score. If the initial face confidence score given by the face feature extraction module is <0.8 or the feature map lacks key points, it is uploaded to the face detection module. If the initial face confidence score given by the face feature extraction module is ≥0.8, it is compared with all face features in the local index. If the score is greater than or equal to 0.75, it is considered to pass. If it is <0.75, it is still uploaded to the face detection module. E3: The face feature extraction module performs a fine inspection of the feature map. If the initial face confidence score is ≥0.9, it is determined to be a credible face. If the initial face confidence score is <0.9, it is considered a mismatch and is sent back to the image integration unit for reshooting.

[0015] A second aspect of the present invention provides a multi-functional standardized acquisition device for railway engineering image data, comprising a multi-functional standardized acquisition system for railway engineering image data and an operating platform. The multi-functional standardized acquisition system for railway engineering image data is any one of the acquisition systems described above, and the multi-functional standardized acquisition system for railway engineering image data is wirelessly connected to the operating platform. Compared with the prior art, the beneficial effects of the present invention are as follows: This application discloses a multi-functional standardized acquisition system and device for railway engineering image data. Through parallel processing between multiple units, it achieves standardized acquisition of railway engineering image data. Features include shooting parameter reminders, voice prompts, OCR pre-recognition, face recognition blurring, image file naming, and video signing. It also supports the extraction of key engineering elements, ensuring the efficiency, accuracy, and security of image data acquisition and guaranteeing the smooth execution of each step while meeting regulatory requirements. Furthermore, the synchronous OCR recognition of signage text, facial feature extraction and verification, further improves data accuracy and consistency, ensuring... Real-time updates and verification of on-site information strengthen the authentication of personnel and equipment, avoid human error and data distortion, and improve the overall reliability and security of the system. Finally, through regular time calibration, the text recognized by OCR is reviewed and high-precision facial recognition is performed to unify the authentication of equipment and personnel, thus ensuring the system's time synchronization, text accuracy, and the rigor of identity authentication, improving the overall accuracy and security of the system, preventing potential vulnerabilities or identity theft, and ensuring the standardized management of the construction site and the authenticity of information. The system is intelligent and lightweight, ensuring accurate analysis and response speed. Attached Figure Description

[0016] Figure 1 System diagram of the multi-functional standardized acquisition system for railway engineering image data provided by the present invention; Figure 2 The overall flowchart for the multi-functional standardized acquisition of railway engineering image data provided by this invention; Figure 3 A flowchart for consistency verification confidence is provided for this invention; Figure 4 This invention provides a cloud-edge-device architecture interaction diagram. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Please see Figures 1-3 As shown in the figure, this embodiment discloses a multi-functional standardized acquisition system for railway engineering image data. The acquisition system includes: an image integration unit, a content detection unit, and a data archiving unit connected in sequence, wherein: In this embodiment, the image processing unit is an edge-side acquisition unit, which can be configured with a Snapdragon 778G 5G processor; it has built-in three-axis accelerometer, magnetometer, ambient light sensor, and gyroscope sensor for level / light level alerts, and is also equipped with a 64MP main camera with OIS and an 8MP wide-angle camera, supporting the main camera for capturing details and the wide-angle camera for capturing panoramas; it has built-in dual-microphone noise reduction, and in addition, to better acquire image data, it is also equipped with a 16kHz mono microphone for voice clarity alerts; it has a 5000mAh + 33W fast charging power supply to meet 8 hours of continuous operation; and it has an IP68 protection rating, withstanding drops of up to 1.5m.

[0019] In addition, the image integration unit can also adopt a 5G / 4G Android rugged handheld device; equipped with SOCDimensity700 / 900 (built-in APU1.2TOPS); the camera uses a rear 50MP PDAF + front 16MP (with LED fill light); USB-C 3.1 interface (for external wide-angle / macro lenses); the APP integrates face detection, text and other models, face recognition uses the MTCNN-light model, text uses the PP-OCRv5_mobile_rec model, sharpness detection and TTS offline voice package for abnormal prompts, and the output data format is JPEG / MP4, JSON. In other implementations, different configurations can be selected according to different situations. In this embodiment, no limitation is made. The effect of image data compression can be seen in Table 1: Table 1: Compression Effect Table ; Simultaneously, the model undergoes secondary compression. Due to the sparse weights resulting in low entropy, the ZIP acts as a container, and the LZMA2 algorithm compresses the weights again. The LZ77 dictionary sliding window (maximum 1GB) replaces repeating byte sequences with "distance-length" pairs; The range encoder (an arithmetic coding variant) performs entropy coding on the remaining symbols after matching based on probability; LZ77 will find tens of thousands of strings of 0 with lengths of 16 / 32 / 64, and store only one "distance-length" string; Resources are compressed: 32 PNG images of the launcher and bootloader are converted to WebP; all icons are converted to SVG. Sample images and help animations are captured and fetched by CDN on demand during first use; subsequent accesses retrieve them directly from the cache node, reducing pressure on the origin server and improving loading speed. APK file paths are configured.

[0020] For example, resource compression first writes all photos or videos into temporary sub-streams (320×240JPEG≤60kB), which are then deleted immediately after the algorithm finishes running; the main stream is a 4K file, which is directly transmitted to the content detection unit via ChunkedStreaming+HTTPS; the cache directory / cache / is set to a quota of 200MB, and when the limit is exceeded, the sub-streams and logs are cleared using LRU.

[0021] Meanwhile, the content detection unit, also known as the edge unit, utilizes an RK3588SoC (8nm, 4×A76+4×A55, 6TOPSNPU) industrial PC for local OCR, 128-dimensional facial feature analysis, and RTC. It is equipped with a time synchronization module supporting both BeiDou and GPS modes, PPS serial port output, and an accuracy range of ±20ns. A temperature-compensated crystal oscillator (TCXO) ensures RTC drift <0.5s / 24h. A 0.33F / 5.5V supercapacitor is included to support RTC maintenance for at least 7 days. The network layer features dual 4G / 5G gate modules and Gigabit Ethernet redundancy. It is IP65 aluminum shell protected, suitable for both tunnel and outdoor use.

[0022] Furthermore, compared to the original MTCNN, MTCNN-Light compromises on accuracy, especially when dealing with side-view or occluded faces, where the false detection rate may be higher. This stems from the reduced feature extraction capability due to model simplification. This paper proposes an improved method that allows MTCNN-Light to maintain its lightweight nature while retaining good accuracy in complex environments. Depthwise separable convolutions can be used to replace ordinary convolutions, reducing the computational cost to 1 / 8 of the original. 1 / 91. Add a lightweight attention mechanism (such as the SE module) to enhance the weights of key features. Formula: SE(x) = x σ(W2δ(W1GAP(x))) where GAP is global average pooling, deltaδ is ReLU, sigmaσ is Sigmoid, and W1 / W2 are the weights of the fully connected layer.

[0023] In this embodiment, for a more intuitive understanding of its functions, please refer to Table 2: Table 2: Function Table ; Meanwhile, the industrial control computer can be placed in the field rack, and the content detection unit can also use NVIDIA Jetson Orin Nano 8GB, Orin Nano 40 TOPS GPU, 256GB storage and 4×PoE gigabit network ports; Ubuntu 20.04 and JetPack 5.1 embedded operating system; the face detection model adopts EdgeFace-S, which compares faces using a project-level 1:N small database; the text detection adopts PP-OCRv5_server_rec, which can perform secondary precision detection of handwritten text, as well as video compression and data caching.

[0024] In this embodiment, the data archiving unit adopts an Ubuntu 22.04 + Kubernetes + Docker system architecture; face recognition uses the FaceNet model; text detection uses PaddleOCR-Server, which supports the recognition of curved text; line violation detection uses SlowFast behavior recognition; and PostgreSQL database and PostGIS spatiotemporal database are used for metadata management. MinIO is used for file object storage.

[0025] Meanwhile, the data archiving unit also serves as the cloud control center, and can also be equipped with two Intel Xeon Gold 6338 rack servers, in a primary / backup active-active configuration, forming a Kubernetes cluster. It features 8 NVIDIA A100 80GB GPUs for Retina Face-ResNet101 and Swin-B+PARSeq high-precision OCR models; a 240TB SASSSD (RAID6) storage solution; 10 Gigabit dual uplink and Gigabit management interfaces per node in the network layer for redundant switching; one BeiDou / GPS dual-mode timing server; a Stratum-1NTP server; and one UPS.

[0026] For its functional data table, please see Table 3: Table 3: Functional Data Table ; S101: Image Integration Unit: Used to assist in guiding the acquisition of images and sound sources, perform preliminary face processing and text recognition on the acquired images and output initial field-level confidence and initial face confidence, as well as pre-name and sign the images; It should be noted that the text on the signs and faces in the images are preprocessed, including image acquisition, perspective correction to eliminate shooting angle deviation, and adaptive histogram equalization for lighting and color. Then, noise reduction, binarization, and normalization are performed.

[0027] Meanwhile, to ensure that the text format of the signage conforms to the standard, the characters are first read out using a model, and those with a confidence level of ≥90% are considered readable. Then, regular expression validation and uniqueness comparison are performed respectively. If both pass, the signage is deemed to conform to the standard.

[0028] It should be understood that the image integration unit includes: a shooting reminder module, an auxiliary guidance module, a clear voice reminder module, a text and face preprocessing module, and a pre-naming module, among which: Shooting reminder module: This module integrates a light sensor, a magnetic field sensor, and an accelerometer. It combines these with a direction rotation matrix function to convert sensor data into direction information, which is then used to detect whether the image is tilted. In this embodiment, image frames are continuously captured via the camera2API interface, supporting real-time preview. Histograms are used to obtain the contrast values ​​of the image, and automatic fill light is applied when the highlight ratio is greater than 70%. AF detection is used for focus. The sensor manager utilizes three types of sensors: type_light, type_accelerometer, and type_magnetic_field. `type_light` obtains the ambient light illuminance (lux); if `lux < 80`, a "low light" message is displayed. `type_accelerometer` obtains the device's linear acceleration along the x, y, and z axes, and `type_magnetic_field` obtains the strength of the surrounding magnetic field along these three axes. Combined with the acceleration, the sensor data is converted into orientation information in a coordinate system using the orientation rotation matrix function `SensorManager.getRotationMatrix` to obtain a more accurate device orientation. This is used to detect whether the photo is tilted. When the absolute value of `roll` is greater than 5°, a "Please keep level" message is displayed.

[0029] During shooting, the camera preview stream detects four QR anchor points in real time, triggering a photo capture every 500ms. A six-tuple dropdown list automatically pops up (item code, inspection batch number, photographer, bystander, unit project, acceptance location). The content detection unit's industrial control computer calls the sign to identify associated information. The T5-Chinese-1.8M micro-generation model, based on the T5 (Text-to-TextTransferTransformer) architecture, infers missing items using identified fields and then uses Redis and regular expressions for secondary validation. A compliant JSON and JWT are generated and uploaded; otherwise, the photo is retaken.

[0030] like Figure 2As shown, the auxiliary guidance module deploys the YOLO-nano-Steel detection model to detect the presence of measuring instruments. If no steel ruler or measuring tape is detected, it prompts "Place the steel ruler in the frame." It calls OCR to check digital sharpness; if the confidence level is less than 0.8, it prompts "Please move closer," and the process takes only 30ms. The algorithm is implemented on the sub-stream.

[0031] Clear Voice Alert Module: It captures the sound source through a microphone and uses noise reduction technology and PicoTTS voice engine to perform linear quantization to obtain the voice clarity. The specific evaluation indicators for speech clarity are: STOI (Short-Term Objective Intelligibility) ≥ 0.75, STI (Speed ​​Transmission Index) ≥ 0.6, and SNR (Signal-to-Noise Ratio) ≥ 15dB. A comprehensive index is obtained based on the above three indicators, and the comprehensive index is ≥ 70.

[0032] Text and Face Preprocessing Module: This module processes images, generates cropped images and converts them into grayscale images, then performs text detection and pre-recognition on the grayscale images, and filters faces within the images in real time. The specific operation involves real-time detection of four QR anchor points in the camera preview stream, triggering the algorithm after 500ms of stabilization. The QR center coordinates, combined with a fixed template for the signage, are used to crop out seven specific regions. Each ROI is uniformly scaled to a 320×32 grayscale image as input. The backbone network uses the lightweight MobileNetV3 convolutional neural network as the feature extraction module for the OCR model, while DBNet enables adaptive binarization for text detection. The CRNN-Lite convolutional recurrent neural network is used as the Chinese character recognition module of the OCR model, outputting {field name, text, confidence score, and character-level confidence vector}.

[0033] Meanwhile, the text-to-face preprocessing module uses MTCNN-Light (1M) for fast filtering to reduce power consumption, achieving a 0.15GFLOP level, and maintaining the same frame rate as the camera for zero latency. Input: Camera preview stream 320×240 sub-stream, data format: RAWNV21byte, output: POST / edge / v1 / face.

[0034] Pre-naming module: Obtains naming rules, pre-names images, and performs quick self-checks using local regular expressions and enumeration values; The specific operation is carried out through the MQTT protocol, which sends rule updates on a daily schedule. In order to ensure that the naming rules are read in time when the server is powered on, the device server needs to be set to hot reload on startup and use MQTT subscription to reuse the cached rules. When the network is interrupted, the rules are added to the local queue and retransmitted after the network is restored.

[0035] As a specific implementation method, the voice clarity reminder module: captures the sound source through a microphone, and uses noise reduction technology and the PicoTTS voice engine to perform linear quantization to obtain the voice clarity. The specific operation is as follows: A1: Real-time audio data acquisition via microphone at a sampling rate of 16kHz and in mono mode; A2: Real-time noise reduction is achieved by using a recurrent neural network (RNN) and combined with optical frequency segmented filtering. The RGB three channels are decomposed, and then guided filtering (GF) is used to denoise and reconstruct each channel separately. This can suppress high-frequency noise in low-light environments while preserving brightness enhancement. A3: Further processing using a 1D-CNN convolutional neural network extracts 320 sampling points, outputs a single float, and then performs linear quantization to obtain speech clarity. Specifically: During the linear quantization process, a quantization scoring threshold is set, and the speech clarity is compared with the quantization scoring threshold; If the speech clarity is less than 50, it is judged as unclear, and a prompt to repeat is made; If 50 ≤ speech clarity < 70, it is judged as unclear, and a prompt is made to repeat it closer; If 70 ≤ speech clarity, the condition is passed, and silence is maintained.

[0036] S102: Content Detection Unit: Used to receive low-confidence data output by the image integration unit, perform secondary recognition of sign text within the image, extract facial features, and perform regular expression verification based on naming. The content detection unit includes: a secondary text recognition module, a facial feature extraction module, and a naming verification module, wherein: The secondary text recognition module uses RESTful and MQTT protocols for transmission and reception. Its dual-channel design facilitates load balancing and fault switching. DBNet++, an improved model based on DBNet, is chosen as the text detection module for secondary OCR verification. It incorporates an Adaptive Scale Fusion (ASF) structure to enhance robustness to curved text. SVTR-Tiny serves as the text recognition module for secondary verification of the OCR model. Content detection units require a fusion confidence score greater than 0.98 to pass; otherwise, the data is passed to the high-precision OCR unit in JSON format with a 112×112 face main stream for secondary text recognition after pre-recognition.

[0037] Face Feature Extraction Module: Detects key facial features within images and extracts feature maps. It uses RetinaFace-ResNet50 (23MB), with a backbone network architecture of ResNet-50+FPN (P3-P7 layers) to improve accuracy (wider-faceHardAP ≈ 90%). Head detection can be performed in parallel across three branches: cls, box, and landmarks, outputting the coordinates of five key points: left eye, right eye, nose, left corner of mouth, and right corner of mouth. 23MB model, 512kb local index, GPU usage <60%. Can run independently even offline. The output five key point format is consistent with MTCNN-Light, allowing for shared affine alignment code. Response time within 1 second. Res50+FPN: WIDER-FaceHardAP 90% (18% better than MTCNN); parallel output of cls+box+5 landmarks, completing detection and alignment in a single forward pass. Output: JSON data format, 112×112 face main stream + 128-dimensional features (cloud side interface) POST / cloud / v1 / face_review; Simultaneously, the EdgeFace-S model was selected to extract five points: left eye, right eye, nose tip, left corner of mouth, and right corner of mouth, from which the affine matrix can be calculated. The angle between the two eyes was calculated, and a 112×112 alignment map was obtained using a rotation matrix. Global texture features were extracted; since the last layer of the EdgeFace-S model network has a fixed output of 512, a 512-dimensional vector was obtained. Further normalization and optimization were then performed.

[0038] Naming validation module: After being uploaded from the pre-naming module to the naming validation module, it performs deduplication again using local regular expressions and enumeration values, generates a compressed file, and uploads it. If the detected file already exists, it immediately returns to rename it. Then, the file is stored locally, and a .JSON file is generated in the same directory as the metadata. After compressing the file and generating a JWT, it continues uploading, supporting resuming interrupted uploads.

[0039] S103: Data Archiving Unit: Used for final verification and archiving of low-confidence data processed by the content inspection unit, as well as for self-verification of identity and unified identity source. As a specific implementation method, the data archiving unit includes: a naming check module, a content verification module, and a face detection module, wherein: Naming check module: Calls the POST / api / v1 / media / upload interface to receive compressed files, adds a trust mark and then stores them in the database. The total time taken is ≤50ms. It should be noted that the multi-module collaboration mechanism is as follows: A POST upload will only be initiated if the regular expression and enumeration pass on the first attempt and the network connection is normal. The naming verification module stores the full enumeration library of daily MQTT pushes, so the content detection unit can match more valid names with the same regular expression. In case of potential inconsistencies, such as files not found in the cache being mistakenly blocked, the naming verification module, upon receiving the file, performs a second regular expression and enumeration using the full library. If it still doesn't match, it returns a 400 Bad Request, and the pre-naming module immediately displays a "format mismatch" message upon receiving it. Before uploading a file, the naming verification module generates a JWT (Lightweight Digital Token), signs it with its private key, and then verifies the signature using its public key. If the JWT is parsed correctly, the signature verification is successful, the file is stored, and its metadata is inserted into the archive.

[0040] The advantage of this step lies in the three-level fusion F=0.9498 achieved through pre-naming module 0.90 filtering, naming verification module 0.98 secondary detection, and naming inspection module 0.995 final inspection, which is 12% better than the single model. Curved / Occluded Text: The DBNet++ ASF structure reduces E2E-CR by 37% (68%→31%), according to data from the internal report "DBNet++ Railway Sign Curved Text Evaluation".

[0041] Content validation module: Further re-examines questionable cropped images, outputs text sequences, and performs regular expression validation on dates; Face detection module: Enhances the fitting ability of irregularly shaped objects, introduces weight coefficients, suppresses weakly correlated background information, and verifies faces within the image.

[0042] The specific implementation is as follows: RetinaFace-ResNet101 is adopted, with the backbone network consisting of ResNet-101 + FPN + deformable convolution (DCNv2). Low-confidence samples in the content detection unit are re-detected on the cloud side, while high-confidence samples are directly included in the sample library. ResNet-101 is a 101-layer residual network, stacking [3, 4, 23, 3] bottleneck residual blocks sequentially, outputting four-level feature maps c2-c5 with strides of 4, 8, 16, and 32, respectively, increasing the receptive field. FPN (Feature Pyramid Network) is fused, generating p3-p7 from top to bottom on top of c2-c5, with strides of 8, 16, 32, 64, and 128, forming a five-layer feature pyramid structure model. Deformable convolution DCNv2 is introduced, which improves the fitting ability for irregularly shaped objects and introduces weight coefficients. mn suppresses irrelevant background information, reducing interference from irrelevant content during feature learning. ResNet-101+FPN+Deformable Convolution (DCNv2) improves the robustness of pose recognition. By combining ResNet-101, FPN, and DCNv2, the detection and keypoint accuracy are significantly better than fixed convolutional networks in extreme scenarios such as lighting conditions, occlusion, small faces, and side profiles in engineering settings.

[0043] Content validation module: Questionable cropped images are re-examined using Swin-DBNet, taking 120ms. Swin-B+PARSeq is used as the core model for high-precision OCR verification. After input image preprocessing, the Swin-B backbone network is integrated with the PARSeq module. Features extracted by Swin-B are input into the PARSeq module, outputting a text sequence. Dates are validated using regularization, with inference time approximately 550ms. fusion confidence scores ≥0.995 are directly accepted; scores <0.995 are labeled and added to the training pool. Manual annotation feedback is initiated for model fine-tuning and distillation. After structured encapsulation, an HTTP request is initiated to transmit the results to the server. Upon receiving the response, the server parses the JSON result, and subsequent uploads are not repeated. The transmitted data format is JSON. Specific operations are as follows: B1: Obtain the initial field-level confidence scores of each module, and assign weights to each module using a static weighting strategy. The weight of the text-face preprocessing module is 0, the weight of the secondary text recognition module is 0.7, and the weight of the content verification module is 0.3. B2: The field-level confidence score of the text is calculated based on the weights of each module, and the expression is as follows: F = 0.7 × se + 0.3 × sc; In the formula: se is the initial field-level confidence score output by the secondary text recognition module, and sc is the field-level confidence score output by the content verification module; B3: Set the threshold for determining the confidence level of the merged field: F≥0.95 is acceptable; F<0.95 triggers manual intervention.

[0044] As a concrete example, when seven fixed ROIs are cut out using QR anchors, ROI number 3 is the "Inspection Date" area.

[0045] If CRNN-Lite yields a score of 0.88, which is lower than 0.90, then the masked image and the original image are packaged together and uploaded.

[0046] A second test was performed on the same ROI (ROI #3), and the SVTR-Tiny test yielded a confidence level of 0.93, which is still lower than 0.98. Therefore, it was marked as "questionable" and uploaded again.

[0047] A high-precision verification was performed, still using ROI #3. PARSeq gave a score of 0.996. This ROI now has two valid scores: se=0.93 (secondary text recognition module); sc=0.996 (Content validation module); Based on the formula for fusion calculation, F = 0.70 × 0.93 + 0.30 × 0.996 ≈ 0.9498; therefore, the final confidence level of this "inspection date" field is 0.9498.

[0048] In the above implementation, the process and collaboration mechanism for obtaining the initial field-level confidence scores of each module are as follows: C1: Four QR codes are detected in real time using a lightweight convolutional recurrent neural network (CRNN-Lite) and QR anchors. After stabilization for 500ms, ROI cropping is triggered. Then, seven fixed regions (ROIs) are cropped according to the template, and each ROI is scaled to a 320×32 grayscale image. The MobileNetV3+DBNet+CRNN-Lite model outputs: field name, text, confidence score, and character-level confidence vector, thereby evaluating the cropped image to obtain the initial field-level confidence score of the text face preprocessing module. If the initial field-level confidence score of the text-face preprocessing module is >0.9, it can pass; if the initial field-level confidence score of the text-face preprocessing module is <0.9, it will be uploaded to the secondary text recognition module. C2: The secondary text recognition module re-detects and recognizes the cropped image and outputs the initial field-level confidence score. If the output initial field-level confidence score is ≥0.98, it passes and is not uploaded to the content verification module. If the output initial field-level confidence score is <0.98, it is uploaded to the content verification module. C3: The content verification module performs accuracy checks on questionable cropped images, provides an initial field-level confidence score, and calculates the merged field-level confidence score.

[0049] The robustness of the model was evaluated using both PR curves and mAP. The targetability score was set to 0.95, and the confidence level was calculated by multiplying the targetability score by the class probability. This primarily measures the probability that a specific class of object actually exists within a bounding box. It is typically obtained by combining the assessment of object existence with the assessment of class classification.

[0050] Face detection module: Enhances the fitting ability for irregularly shaped objects, introduces weighting coefficients, suppresses weakly correlated background information, and verifies faces within the image. Specific operations are as follows: D1: Obtain the initial face confidence scores of each module. The same static weighting strategy is used to assign weights to each module. The weight of the text face preprocessing module is 0, the weight of the face feature extraction module is 0.7, and the weight of the face detection module is 0.3. D2: Calculated based on the initial face confidence scores of each module, expressed as: F = 0.70 × se + 0.30 × sc; In the formula: se is the highest cosine similarity obtained by the face feature extraction module compared with the local index; sc represents the highest cosine similarity obtained by the face detection module compared with the global face database. The process of obtaining the initial face confidence scores for each module is as follows: E1: If the initial confidence level output by the text-face preprocessing module is <0.7, it is discarded and the upload is abandoned. If the initial confidence level output by the text-face preprocessing module is ≥0.7, it is uploaded to the face feature extraction module. E2: The face feature extraction module detects the extracted feature map to obtain an initial face confidence score. If the initial face confidence score given by the face feature extraction module is <0.8 or the feature map lacks key points, it is uploaded to the face detection module. If the initial face confidence score given by the face feature extraction module is ≥0.8, it is compared with all face features in the local index. If the score is greater than or equal to 0.75, it is considered to pass. If it is <0.75, it is still uploaded to the face detection module. If the comparison is successful, the data is written to the database and no longer uploaded. Simultaneously, it is sent back to the image integration unit via MQTT.

[0051] In the event of a network outage, the industrial control computer goes offline, but secondary detection and local comparison remain the same as when online. Each record is written to the local database (SQLite) in real time to ensure no data loss during power outages. Once the network is restored, the industrial control computer compresses and packages the records written to disk in batches and uploads them via MQTT / HTTPS.

[0052] E3: The face feature extraction module performs a fine inspection of the feature map. If the initial face confidence score is ≥0.9, it is determined to be a credible face. If the initial face confidence score is <0.9, it is considered a mismatch and is sent back to the image integration unit for reshooting.

[0053] The comparison rule extracts the embedding vectors and then calculates the Euclidean distance between the two vectors, which reflects the similarity between faces.

[0054] For two embedding vectors v1 and v2, the distance is: d(v1,v2) = ; Where d is the vector dimension, and v1i and v2i are the vector components.

[0055] The decision rule is based on the calculated distance to determine whether two faces belong to the same person; If d(v1, v2) ≤ (where d is the threshold), then it is determined that they are the same person. If d(v1, v2) > 1, then they are different people.

[0056] When cls=0.90, TPR=98% and FPR=1%. Adjusting it further will cause a sharp increase in false alarms. Therefore, 0.90 is set as the "trustworthy face" cutoff point.

[0057] A cos_sim value of 0.82 represents a balance between safety and pass rate (based on internal testing experience, with no publicly published data reports). Therefore, a cos_sim value ≥ 0.82 results in the lowest false recognition rate. If both cls_score ≥ 0.90 and cos_sim ≥ 0.82, the identity is confirmed and considered passed. The sample is added to the "positive sample pool." If cls_score ≥ 0.90 but cos_sim < 0.82, the face is confirmed to be of acceptable quality, but a match was not successful. Further verification is then performed, and the crop image is added to the "manual annotation pool."

[0058] Compared to traditional solutions, it boasts three levels of cascaded accuracy: through triggering, fine-tuning, and final checking, it achieves a 13.6% improvement over a single model (83.2%). Furthermore, due to the integration of the WIDER-FaceHard set, it improves accuracy even in extreme scenarios (side profile / occlusion / small face). Compared to the traditional MTCNN (AP 76.2%), this solution using cloud-based R101+DCNv2 achieves an AP of 92.1% (a 15.9% improvement). Data source: internal report "RetinaFace-ResNet101-DCNv2 Railway Scene Evaluation".

[0059] As a specific implementation method, to verify the consistency of image data content, during image data acquisition, it is linked with the "voice prompt function." The "brief introduction" is converted into text, and then semantic association verification is performed with the image data. The same system clock is used, with an error of <100ms. Resuming interrupted uploads is supported. The converted original voice text is configured as JSON data in sidecar mode for uploading.

[0060] Subsequent lightweight verification involves receiving the file and timestamp, verifying file integrity using MD5 hash values, configuring the RK3588 NPU, embedding the Whisper-Tiny model, and completing speech recognition and text conversion within 120ms. The DBNet++ module is reused to detect all text boxes (including curved ones), outputting polygons (region probability maps) and confidence scores. For high-confidence output images (≥0.98), the SVTR-Tiny module is directly reused to recognize the text; keywords extracted from the speech are paired with the text read from the signboard OCR, and a score is calculated. A score ≥60 is acceptable. A score <60 prompts for re-encoding. Low-confidence output images (<0.98) are uploaded again for high-precision OCR.

[0061] Low-confidence output images (<0.98) are tagged with "Low_conf" and received, with an ACK sent and marked. The Swin-DBNet module is reused to re-examine text regions and output new confidence scores; Swin-B+PARSeq is reused to output text sequences and character-level confidence scores. Regular expression validation is performed on dates in the speech, dates on signs, and actual shooting dates. Images with a fusion confidence score ≥0.995 are directly accepted. The judgment results are then sent back in JSON format via HTTP, updated and marked, without being uploaded repeatedly.

[0062] The above method is based on a gradient strategy: ≥0.995: Archive directly; 0.95–0.994: Data archiving unit manual rapid review (seconds), if correct, archive; if incorrect, notify for re-photographing; <0.95: Immediately notify for a reshoot; The threshold setting criteria are then modified accordingly based on industry-standard business rules: Image Integration Unit 0.9: Lightweight model, prioritizing speed, with moderately relaxed accuracy; Content detection unit 0.98: Medium model, improved accuracy, used as a verification node; Data archiving unit 0.995: Large model, high precision requirements, used as the final decision.

[0063] In addition, to achieve time synchronization, the data archiving unit deploys NTP as a time source, automatically synchronizing daily and periodically down-synchronizing. The content detection unit deploys a secondary time source, receiving time synchronization packets from the data archiving unit via the MQTT protocol. Its local RTC maintains long-term time stability while setting a clear RTC drift limit (±1s / 24h), allowing it to independently maintain its local time during network outages. If the image integration unit's network outage exceeds the threshold, shooting is prohibited. The image integration unit software forcibly synchronizes its time with the content detection unit's time before each startup and shooting session to ensure timestamp consistency. If the image integration unit loses network access during shooting and cannot connect to the content detection unit, it uses the locally cached last synchronized time and requests a timestamp marker. After network recovery, the image integration unit automatically retransmits the "delayed synchronization" marker to the content detection unit. Simultaneously, the system log records the time source, local RTC, and shooting timestamp for auditing purposes.

[0064] The network outage fault tolerance mechanism uses the last cached edge_unix in the image integration unit and the last successfully queried content detection unit time as T_capture; the cache time is marked as not real-time: "time_source": "local_cache" in the sidecar; within 0.5 seconds after the network is restored, the file + sidecar is re-transmitted to the content detection unit, and the content detection unit then gives the flag delay_sync=true to the data archiving unit. If the content detection unit's RTC drift exceeds 1 second / 24 hours, immediately suspend timing synchronization with the image integration unit; prohibit shooting, and simultaneously record the following in the log: {"event":"rtc_fail","drift_sec":1.18}. At the same time, upload the data to the archiving unit. If the data archiving unit detects that T_capture > cloud time + 5s, it will reject the entire packet and return a 400 TIME_FUTURE error to the content inspection unit. The content inspection unit will then mark the task and notify the integration unit via MQTT that the time is abnormal.

[0065] In this embodiment, the encryption method used when uploading images is AES-256-CBC encryption algorithm and mode to protect data security.

[0066] A three-tiered fallback mechanism is also embedded. The overall principle is that if the content detection unit fails, it is cached locally, and then the data archiving unit serves as a fallback for manual review. When the recognition confidence is low, it is marked as low_conf, and the original image plus features are immediately uploaded for a second fine-tuning check in the data archiving unit. If the network is interrupted and the connection drops for more than 30 seconds, it is marked as OFFLINE. Local circular caching and heartbeat retries are used, and data is uploaded in batches after the network is restored to avoid data loss in case of processing failures.

[0067] To prevent duplicate uploads, a unique constraint is imposed on columns or combinations of columns to ensure that there is no duplicate data in the database. The blockchain hash chain is used to generate a fixed-length hash value for each block through a hash function. Different content generates different hash values, and duplicate content can be quickly identified by comparison.

[0068] Time synchronization is achieved by setting an NTP master clock in the data archiving unit, a secondary time source in the edge industrial control computer, and a tertiary time source in the handheld terminal.

[0069] For example, the specific operation of consistency verification is as follows: fragmented CRC check, and each fragment is checked when uploading large files in fragments; Then there's SHA-256 hash verification to prevent file corruption or tampering during transmission; PostGIS spatiotemporal consistency verification is used to check the consistency of image capture time and location. Electronic signatures prevent post-processing tampering with the images.

[0070] In summary, as Figure 4 As shown, the image integration unit, content detection unit, and data archiving unit constitute a cloud-edge-device architecture. That is, the image integration unit is regarded as the device side, the content detection unit is regarded as the edge side, and the data archiving unit is regarded as the cloud control center. The "cloud-edge-device" three-layer architecture can optimize the computing tasks and resource allocation at different levels while ensuring the system's efficiency and flexibility. It also has the ability to perform parallel inference while maintaining intelligence and portability.

[0071] Furthermore, the parallel inference mechanism enables rapid processing of large amounts of input data and immediate responses under high load conditions, significantly enhancing multitasking capabilities and system stability. This efficient collaborative inference strategy not only reduces processing latency but also ensures that various tasks can be performed smoothly in scenarios with high real-time requirements, thereby optimizing the overall user experience and operational efficiency.

[0072] Example 2: This embodiment provides a multi-functional standardized acquisition device for railway engineering image data, including a multi-functional standardized acquisition system for railway engineering image data and an operating platform. The multi-functional standardized acquisition system for railway engineering image data is any of the acquisition systems described above, and the multi-functional standardized acquisition system for railway engineering image data communicates wirelessly with the operating platform.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0074] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-functional standardized acquisition system for railway engineering image data, characterized in that, The system includes: an image integration unit, a content detection unit, and a data archiving unit connected in sequence, wherein: The image integration unit is used to assist in guiding the acquisition of images and sound sources, perform preliminary face processing and text recognition on the acquired images and output initial field-level confidence and initial face confidence, as well as pre-name and sign the images. The content detection unit is used to receive low-confidence data output by the image integration unit, perform secondary recognition of the sign text in the image, extract facial features, and perform regular expression verification based on the naming. The data archiving unit is used to perform final verification and archiving of low-confidence data processed by the content detection unit, as well as to independently verify identity and unify identity sources.

2. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 1, characterized in that, The image processing unit includes: a shooting reminder module, an auxiliary guidance module, a clear voice reminder module, a text and face preprocessing module, and a pre-naming module, wherein: The shooting reminder module integrates a light sensor, a magnetic field sensor, and an acceleration sensor, and combines them with a direction rotation matrix function to convert sensor data into direction information, thereby detecting whether the image is tilted. The auxiliary guidance module deploys a YOLO-nano-Steel detection model to detect the presence of measuring instruments. The voice clarity reminder module: captures the sound source through a microphone, and uses noise reduction technology and PicoTTS voice engine to perform linear quantization to obtain the voice clarity; The text and face preprocessing module is used to process images, generate cropped images and convert them into grayscale images, then perform text detection and pre-recognition on the grayscale images, and filter faces in the images in real time. The pre-naming module: obtains naming rules, pre-names images, and performs rapid self-checks using local regular expressions and enumeration values.

3. The multi-functional standardized acquisition system for railway engineering image data according to claim 2, characterized in that, The voice clarity reminder module: captures the sound source through a microphone, and uses noise reduction technology and the PicoTTS voice engine to perform linear quantization to obtain voice clarity. The specific operation is as follows: A1: Real-time audio data acquisition via microphone at a sampling rate of 16kHz and in mono mode; A2: Real-time noise reduction using a recurrent neural network (RNN); A3: Further processing is performed using a 1D-CNN convolutional neural network to extract sampling points and perform linear quantization to obtain speech intelligibility; Specifically: During the linear quantization process, a quantization scoring threshold is set, and the speech clarity is compared with the quantization scoring threshold; If the speech clarity is less than 50, it is judged as unclear, and a prompt to repeat is made; If 50 ≤ speech clarity < 70, it is judged as unclear, and a prompt is made to repeat it closer; If 70 ≤ speech clarity, the test is passed; remain silent.

4. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 1, characterized in that, The content detection unit includes: a secondary text recognition module, a facial feature extraction module, and a naming verification module, wherein: The secondary text recognition module introduces an adaptive scale fusion structure to improve robustness and performs secondary recognition on the pre-recognized text. The facial feature extraction module detects key facial features within the image and extracts feature maps; The naming verification module: After the pre-naming module uploads to the naming verification module, it uses local regular expressions and enumeration values ​​to perform deduplication detection again, generates a compressed file, and uploads it.

5. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 1, characterized in that, The data archiving unit includes: a naming check module, a content verification module, and a face detection module, wherein: The naming check module receives compressed files, adds a trust flag, and then stores them in the database. The content verification module further re-examines the questionable cropped images, outputs a text sequence, and performs regular expression verification on the dates. The face detection module enhances the fitting ability for irregularly shaped objects, introduces weighting coefficients, suppresses weakly correlated background information, and verifies faces within the image.

6. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 5, characterized in that, The content verification module further re-examines questionable cropped images, outputs a text sequence, and performs regular expression validation on the dates. The specific operations are as follows: B1: Obtain the initial field-level confidence scores of each module, and assign weights to each module using a static weighting strategy. The weight of the text-face preprocessing module is 0, the weight of the secondary text recognition module is 0.7, and the weight of the content verification module is 0.

3. B2: The field-level confidence score of the text is calculated based on the weights of each module, and the expression is as follows: F = 0.7 × se + 0.3 × sc; In the formula: se is the initial field-level confidence score output by the secondary text recognition module, and sc is the field-level confidence score output by the content verification module; B3: Set the threshold for the confidence level of the merged field: F≥0.95 is acceptable; F<0.95 triggers manual intervention.

7. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 6, characterized in that, The process of obtaining the initial field-level confidence scores for each module is as follows: C1: The initial field-level confidence of the text-face preprocessing module is obtained by evaluating the cropped image using the CRNN-Lite lightweight convolutional recurrent neural network; If the initial field-level confidence score of the text-face preprocessing module is >0.9, it can pass; if the initial field-level confidence score of the text-face preprocessing module is <0.9, it will be uploaded to the secondary text recognition module. C2: The secondary text recognition module re-detects and recognizes the cropped image and outputs the initial field-level confidence score. If the output initial field-level confidence score is ≥0.98, it passes and is not uploaded to the content verification module. If the output initial field-level confidence score is <0.98, it is uploaded to the content verification module. C3: The content verification module performs accuracy checks on questionable cropped images, provides an initial field-level confidence score, and calculates the fused field-level confidence score.

8. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 5, characterized in that, The face detection module enhances the fitting ability for irregularly shaped objects by introducing weighting coefficients, suppressing weakly correlated background information, and verifying faces within the image. The specific operations are as follows: D1: Obtain the initial face confidence scores of each module. The same static weighting strategy is used to assign weights to each module. The weight of the text face preprocessing module is 0, the weight of the face feature extraction module is 0.7, and the weight of the face detection module is 0.

3. D2: Calculated based on the initial face confidence scores of each module, expressed as: F = 0.70 × se + 0.30 × sc; In the formula: se is the highest cosine similarity obtained by the face feature extraction module after comparing with the local index; sc represents the highest cosine similarity obtained by the face detection module after comparing with the global face database.

9. The multi-functional standardized acquisition system for railway engineering imagery data according to claim 8, characterized in that, The process of obtaining the initial face confidence scores of each module is as follows: E1: If the initial confidence level output by the text-face preprocessing module is <0.7, it is discarded and the upload is abandoned. If the initial confidence level output by the text-face preprocessing module is ≥0.7, it is uploaded to the face feature extraction module. E2: The face feature extraction module detects the extracted feature map to obtain the initial face confidence score. If the initial face confidence score given by the face feature extraction module is <0.8 or the feature map lacks key points, it is uploaded to the face detection module. If the initial face confidence score given by the face feature extraction module is ≥0.8, it will be compared with all face features in the local index. If the score is greater than or equal to 0.75, it is considered to pass. If it is <0.75, it will still be uploaded to the face detection module. E3: The face feature extraction module performs a fine inspection of the feature map. If the initial face confidence score is ≥0.9, it is determined to be a credible face. If the initial face confidence score is <0.9, it is considered a mismatch and is sent back to the image integration unit for reshooting.

10. A multi-functional standardized acquisition device for railway engineering image data, characterized in that, The system includes a multi-functional standardized acquisition system for railway engineering image data and an operating platform. The multi-functional standardized acquisition system for railway engineering image data is the acquisition system described in any one of claims 1 to 9. The multi-functional standardized acquisition system for railway engineering image data is wirelessly connected to the operating platform.

Images