A machine vision-based method and system for recognizing handwritten numbers on a part surface
By collecting real industrial surface images and constructing wrinkle deformation samples, and combining the YOLO11 model and ordinate clustering and row sorting, the problem of low accuracy in handwritten number recognition in complex industrial scenarios was solved, and efficient number reconstruction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENGZHOU UNIVERSITY OF LIGHT INDUSTRY
- Filing Date
- 2026-05-22
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies have low accuracy in recognizing handwritten numbers in complex industrial scenarios. The training samples do not adequately cover the deformation features of actual industrial surfaces, and they are poorly adaptable to writing conditions such as tilting and multiple lines. Furthermore, there is a lack of effective methods for reconstructing numbers after a single digit is detected and output.
Real industrial surface images were collected and combined with crumpled paper to construct wrinkled deformation samples to generate an extended training sample set. The YOLO11 target detection model was used for digit detection, and the numbers were reconstructed by combining vertical axis clustering and horizontal axis sorting.
It improves the model's adaptability to complex surface deformation, achieves accurate reconstruction from discrete detection results to complete number strings, and significantly improves recognition accuracy.
Smart Images
Figure CN122392039A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of machine vision technology, and in particular to a method and system for recognizing handwritten numbers on the surface of parts based on machine vision. Background Technology
[0002] Accurate reading of part identification information is fundamental in manufacturing execution, assembly verification, batch traceability, and quality inspection. Currently, automated identification methods such as barcodes, QR codes, Radio Frequency Identification (RFID), and laser marking are relatively mature and widely used in industrial production. However, in scenarios involving equipment failure, temporary recording, high-temperature welding, or rapid on-site marking, operators still tend to directly write part numbers, batch numbers, or inspection marks on the part surface. While this method offers convenience and cost advantages, it presents significant technical obstacles to subsequent automated data collection and digital management.
[0003] With the development of deep learning object detection technology, the YOLO series of algorithms has achieved a good balance between detection speed and recognition accuracy, and has gradually become a common solution in industrial vision tasks. However, when it is directly applied to the recognition of handwritten numbers on parts, the training samples do not cover the deformation features of actual industrial surfaces well, and the adaptability to writing conditions such as tilting and multiple lines is poor. Furthermore, there is a lack of effective number reconstruction methods after the detection output of a single digit, resulting in low recognition accuracy in complex industrial scenarios. Summary of the Invention
[0004] Therefore, it is necessary to provide a machine vision-based method and system for recognizing handwritten serial numbers on the surface of parts to address the aforementioned technical problems.
[0005] The following technical solution is adopted in this specification: This specification provides a machine vision-based method for recognizing handwritten serial numbers on the surface of parts, including: Initial image samples of the part surface containing handwritten digit numbers are collected, and the paper containing handwritten digit numbers is crumpled to obtain wrinkled deformation samples. The wrinkled deformation samples are merged with the initial image samples to generate an expanded training sample set. For each image sample in the expanded training sample set, the bounding box and digit label of each handwritten digit are annotated, and a category label file corresponding to each image sample is generated. The expanded training sample set and its category label file together constitute the object detection training dataset. Data augmentation is performed on the training samples in the object detection training dataset to generate an augmented training dataset; a single digit detector is then trained using the augmented training dataset. The target detection image is numbered and identified by a single trained digit detector, and the digit category and bounding box position coordinates of each target detection digit in the target detection image are output. The data is divided into rows based on the y-coordinate of the bounding box positions of the detected numbers. For each row, the bounding boxes are sorted in ascending order according to their x-coordinate values. An ordered number sequence for that row is generated based on the number category corresponding to the detected numbers. All ordered number sequences from all rows are concatenated according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
[0006] Optionally, the data can be divided into rows based on the ordinate of the bounding box position coordinates of the detected target, including: For each detected target digit, clustering is performed based on the ordinate value in its bounding box position coordinates, and bounding boxes belonging to the same ordinate cluster interval are grouped into the same row.
[0007] Optionally, clustering is performed based on the y-coordinate values in the bounding box location coordinates, dividing bounding boxes belonging to the same y-coordinate cluster interval into the same row, including: The DBSCAN algorithm is used to cluster the ordinate values in the bounding box position coordinates. The ordinate value of each bounding box is used as the input data point, and the preset neighborhood distance threshold and preset minimum neighborhood sample number are used as clustering parameters to output at least one cluster. Each cluster corresponds to a ordinate interval. Bounding boxes belonging to the same cluster are grouped into the same row.
[0008] Optionally, the bounding box is a rectangular box that contains the outer range of a single handwritten digit; the digit label is a category identifier from 0 to 9 that identifies the value of the handwritten digit; the category label file is a record file that stores the position coordinates of all bounding boxes and their corresponding digit labels in the same image sample.
[0009] Optionally, data augmentation includes at least one of rotation transformation, shearing transformation, and perspective transformation.
[0010] Optionally, the single digital detector is the YOLO11 object detection model.
[0011] This specification provides a machine vision-based system for recognizing handwritten serial numbers on the surface of parts, including: Sample acquisition and expansion module: used to acquire initial image samples of part surfaces containing handwritten digit numbers, and to crumple the paper containing handwritten digit numbers to obtain wrinkled deformation samples. The wrinkled deformation samples are then merged with the initial image samples to generate an expanded training sample set. The target annotation module is used to annotate the bounding box and digit label of each handwritten digit for each image sample in the expanded training sample set, and generate the category label file corresponding to each image sample. The expanded training sample set and its category label file together constitute the target detection training dataset. Model training module: Used to augment the training samples in the object detection training dataset to generate an augmented training dataset; and to train a single digit detector using the augmented training dataset. Model recognition module: used to identify the numbers of target detection images using a trained single digit detector, and output the digit category and bounding box position coordinates of each target detection digit in the target detection image; The number reconstruction module is used to divide the bounding box coordinates of the detected numbers into rows, sort the bounding boxes in each row in ascending order according to their x-coordinate values, generate an ordered number sequence for that row based on the number category corresponding to the detected number, and concatenate all the ordered number sequences of all rows according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
[0012] Optionally, the system also includes a visualization module; The visualization module, connected to the number reconstruction module, is used to display the original image, detection boxes, and the recognized number string results.
[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described machine vision-based method for recognizing handwritten numbers on the surface of parts.
[0014] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described machine vision-based method for recognizing handwritten numbers on the surface of parts.
[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This specification provides a machine vision-based method for recognizing handwritten numbers on the surface of parts. Initial image samples of the part surface containing handwritten digits are collected, and wrinkled samples are obtained by crumpling the paper. These samples are then merged to generate an expanded training sample set. Each image sample in the training sample set is labeled with a bounding box and a digit label, generating a category label file to form a target detection training dataset. Data augmentation is performed on the training samples to train a single digit detector. The trained digit detector outputs the digit category and bounding box coordinates for each digit in the target detection image. The image is then divided into rows based on the vertical coordinate, and the bounding boxes within each row are sorted by the horizontal coordinate to generate an ordered digit sequence. Finally, the sequences from each row are concatenated according to the vertical coordinate order to reconstruct the complete digit string.
[0016] This invention effectively solves the problem of insufficient coverage of actual industrial surface deformation features by collecting real industrial surface images and combining them with crumpled paper to construct wrinkled deformation samples. This improves the model's adaptability to complex surface deformations such as wrinkles, reflections, and corrosion, and achieves accurate digit recognition. Furthermore, addressing the lack of effective number reconstruction methods after the output of a single digit detection in existing methods, this invention uses a post-processing strategy of vertical axis clustering and horizontal axis sorting to achieve accurate reconstruction from discrete detection results to complete number strings, thereby significantly improving the accuracy of handwritten number recognition in complex industrial scenarios. Attached Figure Description
[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0018] Figure 1 This specification provides a flowchart illustrating a machine vision-based method for recognizing handwritten serial numbers on the surface of parts. Figure 2 This document contains illustrations of some of the training samples provided. Figure 3 This is a schematic diagram of the YOLO11 network structure provided in this specification; Figure 4 A schematic diagram of a machine vision-based handwritten number recognition system for parts surfaces is provided for this specification. Figure 5 This is a schematic diagram of the graphical user interface provided in this specification; Figure 6 This is a schematic diagram of the graphical user interface output results provided in this manual; Figure 7 This is a schematic diagram of a computer device for implementing a machine vision-based method for recognizing handwritten serial numbers on the surface of parts, as provided in this specification. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0020] In existing technologies, handwritten number recognition mainly employs the following schemes: Traditional optical character recognition methods rely on character segmentation, which fails under conditions of continuous writing or tilted strokes, leading to recognition errors; end-to-end detection methods delegate the sorting task to implicit learning by the network, but the order is easily disrupted when the digit spacing is uneven or multiple lines are arranged; multi-level network methods rely on prior knowledge of the layout, resulting in poor adaptability to changes in line order; traditional feature methods rely on manually designed features, which lack robustness in complex backgrounds. Therefore, the technical bottleneck of existing technologies lies in the insufficient coverage of deformation features of actual industrial surfaces by training samples, poor adaptability to writing conditions such as tilting and multiple lines, and the lack of effective number reconstruction methods after the output of a single digit detection, resulting in low recognition accuracy in complex industrial scenarios.
[0021] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0022] Figure 1 This is a flowchart illustrating a machine vision-based method for recognizing handwritten serial numbers on the surface of parts, as described in this specification. The method includes the following steps: S101: Collect initial image samples of the part surface containing handwritten digit numbers, and crumple the paper containing handwritten digit numbers to obtain wrinkled deformation samples. Merge the wrinkled deformation samples with the initial image samples to generate an expanded training sample set.
[0023] In industrial settings, handwritten serial numbers on parts are subject to interference from varied writing styles, complex backgrounds, localized glare, or dirt. Furthermore, the distribution of samples in existing public datasets differs significantly from that in real industrial settings, resulting in insufficient generalization ability of models trained on public datasets in practical applications.
[0024] Based on this, in one embodiment of this specification, on the one hand, initial image samples of the surface of real industrial parts are acquired, including images of handwritten numbers on reflective surfaces, complex textured surfaces, rusted surfaces, and dusty surfaces; on the other hand, wrinkled deformation samples are constructed by crumpling the paper with written numbers to simulate the character deformation caused by paper bending and uneven surfaces in industrial settings. Figure 2 As shown, Figure 2The following are some of the training sample images provided in this manual: (a) is an image of a handwritten number on a complex textured surface; (b) is an image of a handwritten number on a rusted surface; (c) is an image of a handwritten number on a dusty surface; and (d) is a wrinkled and deformed sample. The expanded training sample set formed by merging the two types of samples can more comprehensively cover the complex surface morphology in industrial scenarios, providing a data foundation for subsequent model training.
[0025] S102: For each image sample in the expanded training sample set, annotate the bounding box and digit label of each handwritten digit, and generate the category label file corresponding to each image sample. The expanded training sample set and its category label file together constitute the object detection training dataset.
[0026] In one embodiment of this specification, the bounding box is a rectangular box that includes the outer range of a single handwritten digit; the digit label is a category identifier from 0 to 9 that identifies the value of the handwritten digit; and the category label file is a record file that stores the position coordinates of all bounding boxes and their corresponding digit labels in the same image sample.
[0027] S103: Perform data augmentation on the training samples in the object detection training dataset to generate an augmented training dataset; use the augmented training dataset to train a single digit detector.
[0028] To address potential tilt and perspective variations in digital displays in industrial settings, existing methods typically rely on tilt correction algorithms during the preprocessing stage, but these methods exhibit poor robustness against complex backgrounds (reflection, corrosion, texture interference).
[0029] In one embodiment of this specification, data augmentation includes at least one of rotation transformation, shearing transformation, and perspective transformation.
[0030] Furthermore, in one embodiment of this specification, by setting the rotation angle to vary randomly within a small range and introducing moderate shearing and slight perspective perturbation, the detection stability of the model under non-orthogonal viewpoint and curved surface conditions is improved.
[0031] S104: The target detection image is numbered and identified by a single trained digit detector, and the digit category and bounding box position coordinates of each target detection digit in the target detection image are output.
[0032] In one embodiment of this specification, the single digital detector is the YOLO11 target detection model.
[0033] The YOLO11 object detection model includes a backbone network, a neck network, and a detection head, such as... Figure 3 As shown, Figure 3This is a schematic diagram of the YOLO11 network structure provided by the present invention. The backbone network extracts multi-scale feature maps of the input image through multi-layer convolution operations, which can simultaneously capture the local details of small-sized digits and the global structure of large-sized digits. The neck network fuses feature information from different levels to enhance the model's adaptability to scale changes. The detection head outputs the bounding box position, confidence level, and digit category based on the fused feature map.
[0034] In one embodiment of this specification, the model is trained based on the training set, and the detection accuracy is evaluated on test data. Experimental results show that the mean accuracy (mAP@0.5) for a single digit detection task with an intersection-union ratio (IUU) threshold of 0.5 can reach 0.955, the accuracy of recognizing complete digit strings under complex image conditions reaches 67.4%, and the average inference time per image is approximately 31.1 milliseconds.
[0035] By using YOLO11 as a single digit detector, the problems of large differences in digit size and easy loss of details or inability to cover global information when extracting features from a fixed receptive field in industrial scenarios are solved, providing reliable detection results for subsequent numbering reconstruction.
[0036] S105: Divide the bounding box coordinates of the detected numbers into rows. Sort the bounding boxes in each row in ascending order according to their x-coordinate values. Generate an ordered number sequence for that row based on the number category corresponding to the detected number. Concatenate all the ordered number sequences in all rows according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
[0037] In one embodiment of this specification, dividing the target detection digits into rows based on the ordinate of the bounding box position coordinates includes: for each target detection digit, performing clustering based on the ordinate value in its bounding box position coordinates, and dividing bounding boxes belonging to the same ordinate clustering interval into the same row.
[0038] Furthermore, in one embodiment of this specification, clustering is performed based on the ordinate values in the bounding box position coordinates to divide bounding boxes belonging to the same ordinate cluster interval into the same row. This includes: using the DBSCAN algorithm to cluster the ordinate values in the bounding box position coordinates, taking the ordinate value of each bounding box as the input data point, using a preset neighborhood distance threshold and a preset minimum number of neighborhood samples as clustering parameters, and outputting at least one cluster; each cluster corresponds to a ordinate interval; and dividing bounding boxes belonging to the same cluster into the same row.
[0039] For each target detection digit, the ordinate of the center point of its bounding box is obtained as the ordinate of the target detection digit. The ordinate values of all target detection digits are used to form a one-dimensional data point set. Clustering is performed on this set, and data points with similar ordinate values are grouped into the same cluster. Each cluster corresponds to a ordinate interval, and the bounding boxes belonging to the same cluster are divided into the same row.
[0040] Clustering can be performed using the DBSCAN algorithm. The input data point is the y-coordinate of each detected digit. Preset neighborhood distance thresholds and a preset minimum number of neighborhood samples are used as clustering parameters, outputting at least one cluster. The neighborhood distance threshold determines whether two y-coordinate values belong to the same neighborhood; they are considered neighbors when the difference between the two y-coordinate values is less than or equal to the threshold. The minimum number of neighborhood samples is used to determine core points; it is set to 1, meaning each row contains at least one digit.
[0041] Through the above line splitting process, the line order of multi-line numbering can be automatically identified without relying on prior knowledge of the layout, and it has good adaptability to uneven line spacing, slanted number arrangement, and other situations.
[0042] The execution subject of the methods provided in this specification can be a server, which can be a server set up on a business platform, or a device such as a desktop computer or laptop computer that can execute the scheme in this specification.
[0043] based on Figure 1 This paper presents a machine vision-based method for recognizing handwritten serial numbers on parts surfaces. By acquiring images of real industrial surfaces and combining them with crumpled paper to construct wrinkled deformation samples, it effectively solves the problem of insufficient coverage of deformation features of actual industrial surfaces by training samples. This improves the model's adaptability to complex surface deformations such as wrinkles, reflections, and corrosion, and achieves accurate digit recognition. Furthermore, addressing the lack of effective number reconstruction methods after the detection output of a single digit in existing methods, this paper adopts a post-processing strategy of vertical axis clustering and horizontal axis sorting to achieve accurate reconstruction from discrete detection results to complete number strings, thereby significantly improving the accuracy of handwritten serial number recognition in complex industrial scenarios.
[0044] When applying the machine vision-based method for recognizing handwritten part numbers provided in this manual, it is not necessary to... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.
[0045] The above describes a machine vision-based method for recognizing handwritten numbers on part surfaces, provided by one or more embodiments of this specification. Based on the same approach, this specification also provides a corresponding machine vision-based system for recognizing handwritten numbers on part surfaces, such as... Figure 4As shown.
[0046] Figure 4 This specification provides a schematic diagram of a machine vision-based part surface handwritten number recognition system, including: Sample acquisition and expansion module: used to acquire initial image samples of part surfaces containing handwritten digit numbers, and to crumple the paper containing handwritten digit numbers to obtain wrinkled deformation samples. The wrinkled deformation samples are then merged with the initial image samples to generate an expanded training sample set. The target annotation module is used to annotate the bounding box and digit label of each handwritten digit for each image sample in the expanded training sample set, and generate the category label file corresponding to each image sample. The expanded training sample set and its category label file together constitute the target detection training dataset. Model training module: Used to augment the training samples in the object detection training dataset to generate an augmented training dataset; and to train a single digit detector using the augmented training dataset. Model recognition module: used to identify the numbers of target detection images using a trained single digit detector, and output the digit category and bounding box position coordinates of each target detection digit in the target detection image; The number reconstruction module is used to divide the bounding box coordinates of the detected numbers into rows, sort the bounding boxes in each row in ascending order according to their x-coordinate values, generate an ordered number sequence for that row based on the number category corresponding to the detected number, and concatenate all the ordered number sequences of all rows according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
[0047] In one embodiment of this specification, the system further includes a visualization module; Visualization modules, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the graphical user interface provided in this manual. It connects to the number reconstruction module and is used to display the original image, detection boxes, and the recognized number string results. The output results are as follows: Figure 6 As shown, Figure 6 This is a schematic diagram of the graphical user interface output provided in this manual.
[0048] For specific limitations regarding a machine vision-based handwritten number recognition system for parts surfaces, please refer to the limitations of a machine vision-based handwritten number recognition method for parts surfaces described above, which will not be repeated here. Each module in the aforementioned machine vision-based handwritten number recognition system for parts surfaces can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0049] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A machine vision-based method for recognizing handwritten serial numbers on the surface of parts is provided.
[0050] This instruction manual also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 A machine vision-based method for recognizing handwritten serial numbers on the surface of parts is provided.
[0051] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A method for recognizing handwritten serial numbers on the surface of parts based on machine vision, characterized in that, include: Initial image samples of the part surface containing handwritten digit numbers are collected, and the paper containing handwritten digit numbers is crumpled to obtain wrinkled deformation samples. The wrinkled deformation samples are merged with the initial image samples to generate an expanded training sample set. For each image sample in the expanded training sample set, the bounding box and digit label of each handwritten digit are annotated, and a category label file corresponding to each image sample is generated. The expanded training sample set and its category label file together constitute the object detection training dataset. Data augmentation is performed on the training samples in the object detection training dataset to generate an augmented training dataset; Train a single digit detector using the enhanced training dataset; The target detection image is numbered and identified by a single trained digit detector, and the digit category and bounding box position coordinates of each target detection digit in the target detection image are output. The data is divided into rows based on the y-coordinate of the bounding box positions of the detected numbers. For each row, the bounding boxes are sorted in ascending order according to their x-coordinate values. An ordered number sequence for that row is generated based on the number category corresponding to the detected numbers. All ordered number sequences from all rows are concatenated according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
2. The method as described in claim 1, characterized in that, The data is divided into rows based on the y-coordinate of the bounding box position coordinates of the detected target, including: For each detected target digit, clustering is performed based on the ordinate value in its bounding box position coordinates, and bounding boxes belonging to the same ordinate cluster interval are grouped into the same row.
3. The method as described in claim 2, characterized in that, Clustering is performed based on the y-coordinate values of the bounding box locations, grouping bounding boxes belonging to the same y-coordinate cluster interval into the same row, including: The DBSCAN algorithm is used to cluster the ordinate values in the bounding box position coordinates. The ordinate value of each bounding box is used as the input data point, and the preset neighborhood distance threshold and preset minimum neighborhood sample number are used as clustering parameters to output at least one cluster. Each cluster corresponds to a ordinate interval. Bounding boxes belonging to the same cluster are grouped into the same row.
4. The method as described in claim 1, characterized in that, The bounding box is a rectangular box that contains the outer range of a single handwritten digit; the digit label is a category label from 0 to 9 that identifies the value of the handwritten digit; the category label file is a record file that stores the position coordinates of all bounding boxes and their corresponding digit labels in the same image sample.
5. The method as described in claim 1, characterized in that, Data augmentation includes at least one of rotation transformation, shearing transformation, and perspective transformation.
6. The method as described in claim 1, characterized in that, The single digital detector is the YOLO11 object detection model.
7. A machine vision-based system for recognizing handwritten serial numbers on the surface of parts, characterized in that, include: Sample acquisition and expansion module: used to acquire initial image samples of part surfaces containing handwritten digit numbers, and to crumple the paper containing handwritten digit numbers to obtain wrinkled deformation samples. The wrinkled deformation samples are then merged with the initial image samples to generate an expanded training sample set. The target annotation module is used to annotate the bounding box and digit label of each handwritten digit for each image sample in the expanded training sample set, and generate the category label file corresponding to each image sample. The expanded training sample set and its category label file together constitute the target detection training dataset. Model training module: Used to perform data augmentation on the training samples in the object detection training dataset and generate an augmented training dataset; Train a single digit detector using the enhanced training dataset; Model recognition module: used to identify the numbers of target detection images using a trained single digit detector, and output the digit category and bounding box position coordinates of each target detection digit in the target detection image; The number reconstruction module is used to divide the bounding box coordinates of the detected numbers into rows, sort the bounding boxes in each row in ascending order according to their x-coordinate values, generate an ordered number sequence for that row based on the number category corresponding to the detected number, and concatenate all the ordered number sequences of all rows according to the numerical order of the clustering intervals of the y-coordinates to reconstruct the complete number string.
8. The system as described in claim 7, characterized in that, The system also includes a visualization module; The visualization module, connected to the number reconstruction module, is used to display the original image, detection boxes, and the recognized number string results.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 6.
10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 6.