An image recognition method and device for informatization archiving of archives

Abstract

The application provides an image recognition method and device for informatization archiving of archives. The method comprises: obtaining printed text content and a plurality of handwritten content area images by performing image recognition on a target image of an archive to be processed; performing clustering analysis on the plurality of handwritten content area images to determine a target classification cluster to which each area image belongs; selecting a target recognition model for the area image from a plurality of lightweight handwritten content recognition models according to an accuracy score corresponding to the target classification cluster; inputting each area image into the corresponding target recognition model to recognize the handwritten content, and obtaining a handwritten content recognition result of each area image; and finally obtaining electronic archiving content of the archive to be processed according to the printed text content and the handwritten content recognition result. The handwritten content can be recognized specifically, thereby significantly improving the accuracy and efficiency of handwritten content recognition.

Description

Technical Field

[0001] This application relates to the technical field of archival information filing, specifically to an image recognition method and apparatus for archival information filing. Background Technology

[0002] Information-based archiving refers to the entire process of collecting, integrating, maintaining, processing, and providing access to archival information resources using modern information technology. This marks a strategic shift in archival management from physical preservation to digital information services. Information-based archiving also involves original paper documents of enterprises. Current technology involves scanning or photographing these original documents to obtain corresponding image data, which is then recognized using handwriting recognition models to convert them into electronic archives, or extracting key content as an index, thereby achieving automated classification, retrieval, and management of archives.

[0003] However, in university professional teaching, industry-education integration practices, and Lingnan culture research, a large amount of paper archives and documents have long needed to be organized, archived, digitized, and preserved long-term. These archives are diverse in form, including standard archives with printed fonts, as well as a large amount of non-standard handwritten content such as handwritten notes, corrections, annotations, and doodles with varying handwriting styles, severe ligatures, and faded paper due to aging. Current recognition models can only accurately identify printed content and cannot accurately identify handwritten content. Therefore, existing archival digitization methods cannot accurately and effectively identify and archive archives containing handwritten content. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings and deficiencies of the existing technology and provide an image recognition method and device for archiving information, which can perform targeted recognition of handwritten content, thereby significantly improving the accuracy and efficiency of handwritten content recognition and thus improving the overall quality of archiving information.

[0005] The first aspect of this application provides an image recognition method for archiving information systems, including: Cluster analysis was performed on image samples containing multiple handwritten text to obtain multiple classification clusters; Image samples of multiple handwritten contents are input into multiple lightweight handwritten content recognition models to obtain handwritten content recognition result samples output by each lightweight handwritten content recognition model, as well as the accuracy score of each lightweight handwritten content recognition model for each classification cluster. Image recognition is performed on the target image of the file to be processed to obtain the printed text content of the target image and the area images of multiple handwritten contents; Cluster analysis is performed on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the multiple classification clusters; Based on the accuracy score corresponding to the target classification cluster, a target recognition model for the region image is selected from the plurality of lightweight handwritten content recognition models. Each of the aforementioned regions is input into the corresponding target recognition model to perform handwritten content recognition, thereby obtaining the handwritten content recognition results for each region image. Based on the printed text content results and the handwritten content recognition results, the electronic archive content of the file to be processed is obtained.

[0006] As one implementation method, the step of selecting a target recognition model for the region image from the plurality of lightweight handwritten content recognition models based on the accuracy score corresponding to the target classification cluster includes: The task queue of the lightweight handwritten content recognition model with the highest accuracy score corresponding to the target classification cluster; If the number of queued tasks in the task queue is less than or equal to a preset threshold, the lightweight handwritten content recognition model with the highest accuracy score will be determined as the target recognition model. If the number of queued tasks in the task queue is greater than the preset threshold, an allocation request is sent to the trained agent model to obtain the target recognition model selected by the agent model.

[0007] As one implementation method, the agent model includes a policy network and a value network; the agent model is trained through the following steps: Based on multiple image samples of handwritten content, generate multiple sample allocation requests and send them continuously and randomly to the initialized agent model; Upon receiving a sample allocation request, the policy network performs a model selection action based on the current environment state, selecting a sample recognition model from among the various lightweight handwritten content recognition models and updating the environment state; wherein, the current environment state includes the current task queue of each lightweight handwritten content recognition model and the accuracy score corresponding to each classification cluster; Based on the environmental conditions, obtain the instantaneous reward for the action selected by the model; The estimated expected value of the selected action is obtained through the value network; The environmental state, model selection action, instantaneous reward, and estimated expected value corresponding to each allocation request sample are stored as training samples in the experience pool. The network parameters of the value network and the policy network are updated based on the training samples stored in the experience pool to obtain the agent model.

[0008] As one implementation method, the step of obtaining the instantaneous reward of the model's selected action based on the environmental state includes: From the current environment state, obtain the sample accuracy score of the sample recognition model for the classification cluster to which the image sample corresponding to the received sample allocation request belongs; If the sample accuracy score is less than a preset score threshold, the instantaneous reward is determined to be a negative reward. If the sample accuracy score is greater than or equal to the preset score threshold, the instantaneous reward for the model's selected action is obtained based on the current task queue of the sample recognition model and the sample accuracy score.

[0009] As one implementation method, the step of obtaining the instantaneous reward for the action selected by the model based on the current task queue of the sample recognition model and the sample accuracy score includes: The instantaneous reward is obtained using the following formula: ; in, For instant reward, The number of tasks in the current task queue of the sample recognition model. Adjust the coefficients for the task. Score the accuracy of the sample. The preset scoring threshold is defined as follows.

[0010] As one implementation method, the step of updating the network parameters of the value network and the policy network based on the training samples stored in the experience pool to obtain the agent model includes: The advantage function is obtained based on the instantaneous reward and estimated expected value of the training samples; The network parameters of the value network and the policy network are updated according to the advantage function.

[0011] As one implementation method, the step of obtaining the advantage function based on the instantaneous reward and estimated expected value of the training samples includes: The advantage function is obtained using the following formula: ; in, For the aforementioned advantage function, For environmental conditions, Select actions for the model. For the network parameters of the policy network, For the network parameters of the value network, The number of training samples, As a discount factor, The instantaneous reward for the corresponding training sample. This is the estimated expected value.

[0012] As one implementation, the step of updating the network parameters of the value network and the policy network according to the advantage function includes: The network parameters of the value network are updated using the following formula: ; in, The updated network parameters of the value network, The network parameters of the value network before the update. The learning rate of the value network. The number of training samples, For partial derivatives, This is the advantage function.

[0013] As one implementation, the step of updating the network parameters of the value network and the policy network according to the advantage function includes: The network parameters of the policy network are updated using the following formula: ; in, The updated network parameters of the policy network, The network parameters of the policy network before the update. The number of training samples, For splitting operators, The output of the policy network, For the aforementioned advantage function, The strength coefficient of the entropy regularization term. Let be the entropy of the policy distribution.

[0014] Compared to related technologies, the image recognition method for archival information archiving in this application involves image recognition of the target image of the archive to be processed, obtaining the printed text content and multiple handwritten content region images of the target image, then performing cluster analysis on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs. Based on the accuracy score corresponding to the target classification cluster, a target recognition model for the region image is selected from multiple lightweight handwritten content recognition models. Each region image is then input into the corresponding target recognition model for handwritten content recognition, obtaining the handwritten content recognition result for each region image. Finally, based on the printed text content result and the handwritten content recognition result, the electronic archival content of the archive to be processed is obtained. This method allows for targeted recognition of handwritten content, significantly improving the accuracy and efficiency of handwritten content recognition, thereby enhancing the overall quality of archival information archiving.

[0015] A second aspect of this application provides an image recognition device for archiving information systems, comprising: The clustering analysis module is used to perform clustering analysis on image samples of multiple handwritten contents to obtain multiple classification clusters; The accuracy score acquisition module is used to input image samples of multiple handwritten contents into multiple lightweight handwritten content recognition models, obtain handwritten content recognition result samples output by each lightweight handwritten content recognition model, and the accuracy score of each lightweight handwritten content recognition model for each classification cluster. The image recognition module is used to perform image recognition on the target image of the file to be processed, and obtain the printed text content of the target image and the area image of multiple handwritten contents; The target classification cluster acquisition module is used to perform cluster analysis on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the multiple classification clusters; The target recognition model acquisition module is used to select the target recognition model of the region image from the plurality of lightweight handwritten content recognition models according to the accuracy score corresponding to the target classification cluster. The handwritten content recognition module is used to input the images of each region into the corresponding target recognition model to perform handwritten content recognition and obtain the handwritten content recognition results of each region image. The electronic archive content acquisition module is used to acquire the electronic archive content of the archive to be processed based on the printed text content result and the handwritten content recognition result.

[0016] Compared to related technologies, the image recognition device for archival information archiving of this application, after performing image recognition on the target image of the archive to be processed to obtain the printed text content result and multiple handwritten content area images of the target image, performs cluster analysis on the multiple handwritten content area images to determine the target classification cluster to which each area image belongs. Then, based on the accuracy score corresponding to the target classification cluster, a target recognition model for the area image is selected from multiple lightweight handwritten content recognition models. Each area image is then input into the corresponding target recognition model for handwritten content recognition to obtain the handwritten content recognition result for each area image. Finally, based on the printed text content result and the handwritten content recognition result, the electronic archival content of the archive to be processed is obtained. This device can perform targeted recognition of handwritten content, thereby significantly improving the accuracy and efficiency of handwritten content recognition, and thus improving the overall quality of archival information archiving.

[0017] To provide a clearer understanding of this application, the specific embodiments of this application will be described below in conjunction with the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a flowchart of an image recognition method for archiving information in an embodiment of this application.

[0019] Figure 2 This is a flowchart of step S5 of an image recognition method for archiving information in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the module connections of an image recognition device for archiving information in an embodiment of this application.

[0021] 100. Image recognition device; 101. Cluster analysis module; 102. Accuracy score acquisition module; 103. Image recognition module; 104. Target classification cluster acquisition module; 105. Target recognition model acquisition module; 106. Handwritten content recognition module; 107. Electronic archive content acquisition module. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0023] It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0024] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. The singular forms "a," "the," and "the" used in this application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. The word "if" as used herein can be interpreted as "when," "when," or "in response to determination."

[0025] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0026] Please see Figure 1 This is a flowchart of an image recognition method for archiving information management according to the first embodiment of this application. The method includes: S1: Cluster analysis is performed on image samples of multiple handwritten contents to obtain multiple classification clusters.

[0027] Image samples of handwritten content include images of the handwritten content and corresponding electronic identifiers. The handwritten content includes, but is not limited to, handwritten text and handwritten symbols, while the electronic identifiers include, but are not limited to, electronic text and electronic symbols.

[0028] Cluster analysis is a typical unsupervised data analysis and machine learning method. Its core principle follows the logic of "grouping by category," requiring no manually labeled data or pre-defined classification rules. It automatically groups samples with similar features into the same cluster and samples with significantly different features into different clusters based solely on the similarity of the data's features. This allows for the discovery of hidden distribution patterns and correlations within the data, achieving the simplification and pattern extraction of messy data. In this embodiment, by clustering the handwritten content features of image samples (such as handwriting style, stroke connections, corrections / annotations, drafts, etc.), multiple clusters corresponding to different handwritten content feature types can be formed. This automatic grouping based on feature similarity not only effectively avoids the subjective bias and high cost of manual classification but also lays the data foundation for subsequently matching the optimal recognition model to different clusters.

[0029] S2: Input multiple handwritten content image samples into multiple lightweight handwritten content recognition models, obtain the handwritten content recognition result samples output by each lightweight handwritten content recognition model, and the accuracy score of each lightweight handwritten content recognition model for each classification cluster.

[0030] The lightweight handwritten content recognition model can be obtained by pruning, quantizing, or knowledge distilling the handwritten content recognition model. The handwritten content recognition model can be a publicly shared network model for recognizing handwritten content, or a deep learning-based network model, such as a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), or a Transformer architecture based on an attention mechanism. Among these multiple lightweight handwritten content recognition models, there are multiple models with different architectures, as well as multiple models with the same architecture but different model parameters. This diversity in model design can cover recognition needs in different scenarios.

[0031] S3: Perform image recognition on the target image of the file to be processed to obtain the printed text content of the target image and the area images of multiple handwritten contents.

[0032] The printed text content can be obtained by recognizing the printed areas in the target image using Optical Character Recognition (OCR) technology, while the handwritten content area image is obtained through object detection or semantic segmentation technology. In some embodiments, areas where the confidence level of Optical Character Recognition (OCR) is less than a preset confidence threshold can be identified as handwritten content areas, thereby extracting more accurate handwritten content area images.

[0033] S4: Perform cluster analysis on the region images of the multiple handwritten contents to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the multiple classification clusters.

[0034] S5: Based on the accuracy score corresponding to the target classification cluster, select the target recognition model for the region image from the multiple lightweight handwritten content recognition models.

[0035] Please see Figure 2 Step S5 includes: S51: The task queue of the lightweight handwritten content recognition model with the highest accuracy score corresponding to the target classification cluster.

[0036] A task queue is a buffer queue used to temporarily store pending identification tasks. Tasks temporarily stored in the task queue are called queued tasks. The more queued tasks in the task queue, the higher the current load on the model or the longer the waiting time. Continuing to assign tasks to it may lead to increased processing latency or response timeouts. Conversely, the fewer queued tasks in the task queue, the more idle resources the model currently has, and the faster the response speed.

[0037] S52: If the number of queued tasks in the task queue is less than or equal to a preset threshold, the lightweight handwritten content recognition model with the highest accuracy score is determined as the target recognition model.

[0038] When the number of tasks in the task queue is less than or equal to a preset threshold, it indicates that the model is under low load and has sufficient computing resources to quickly respond to recognition requests. In this case, directly selecting this model as the target recognition model can achieve the fastest processing speed while ensuring recognition accuracy, which is beneficial for efficiently and accurately obtaining handwritten content recognition results.

[0039] S53: If the number of queued tasks in the task queue is greater than the preset number threshold, an allocation request is sent to the trained agent model to obtain the target recognition model selected by the agent model.

[0040] When the number of queued tasks in the task queue exceeds the preset threshold, it indicates that the model is currently under high load. Forcibly assigning tasks may lead to response delays or resource congestion. To avoid this situation, the target recognition model is selected from multiple lightweight models by the agent model, which can achieve recognition accuracy and load balancing. While ensuring recognition accuracy, it avoids overloading a single model and optimizes the overall recognition efficiency.

[0041] The agent model includes a policy network and a value network; the agent model is trained through the following steps: S5301: Based on multiple image samples of handwritten content, generate multiple sample allocation requests and send them continuously and randomly to the initialized intelligent agent model; S5302: When a sample allocation request is received, the policy network performs a model selection action based on the current environment state, selects a sample recognition model from the various lightweight handwritten content recognition models, and updates the environment state; wherein, the current environment state includes the current task queue of each lightweight handwritten content recognition model and the accuracy score corresponding to each classification cluster; S5303: Based on the environmental state, obtain the instantaneous reward of the action selected by the model.

[0042] Instantaneous rewards are evaluations of the model's selected action based on the environmental state, providing immediate feedback on the quality of the model selection. For example, if the selected sample recognition model has high accuracy relative to the image sample, a positive reward is given; if the selected sample recognition model has low accuracy relative to the image sample, a negative reward is given. In other embodiments, instantaneous rewards can also be adjusted based on the current task queue of the selected sample recognition model. For example, if the current task queue of the selected sample recognition model is large, the instantaneous reward is reduced; if the current task queue is small, the instantaneous reward is increased. Through this dynamic adjustment mechanism, the agent model can learn to pursue high recognition accuracy while maintaining a balanced workload across models.

[0043] S5304: Obtain the estimated expected value of the selected action through the value network.

[0044] Among them, the estimated expected value is the mathematical expectation of the sum of all future discount rewards under the current strategy, based on the environmental conditions and the continuous action. It is used to measure the merits of the chosen actions under the current strategy in the long run.

[0045] S5305: Store the environmental state, model-selected action, instantaneous reward, and estimated expected value corresponding to each allocation request sample as training samples in the experience pool.

[0046] S5306: Update the network parameters of the value network and the policy network according to the training samples stored in the experience pool to obtain the agent model.

[0047] S6: Input the images of each region into the corresponding target recognition model to perform handwritten content recognition, and obtain the handwritten content recognition results of each region image.

[0048] S7: Based on the printed text content result and the handwritten content recognition result, obtain the electronic archive content of the file to be processed.

[0049] The electronic archive content can be keywords extracted from the printed text content results and the handwritten content recognition results, or it can be an electronic document generated based on the printed text content results and the handwritten content recognition results. For example, based on the coordinate region of the handwritten content's regional image in the target image of the archive to be processed, the handwritten content recognition results are combined with the printed text content results to generate an electronic document.

[0050] Compared to related technologies, the image recognition method for archival information archiving in this application involves image recognition of the target image of the archive to be processed, obtaining the printed text content and multiple handwritten content region images of the target image, then performing cluster analysis on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs. Based on the accuracy score corresponding to the target classification cluster, a target recognition model for the region image is selected from multiple lightweight handwritten content recognition models. Each region image is then input into the corresponding target recognition model for handwritten content recognition, obtaining the handwritten content recognition result for each region image. Finally, based on the printed text content result and the handwritten content recognition result, the electronic archival content of the archive to be processed is obtained. This method allows for targeted recognition of handwritten content, significantly improving the accuracy and efficiency of handwritten content recognition, thereby enhancing the overall quality of archival information archiving.

[0051] In a feasible embodiment, S5303: the step of obtaining the instantaneous reward of the model-selected action based on the environmental state includes: S53031: From the current environment state, obtain the sample accuracy score of the sample recognition model for the classification cluster to which the image sample corresponding to the received sample allocation request belongs.

[0052] S53032: If the sample accuracy score is less than the preset score threshold, the instantaneous reward is determined to be a negative reward.

[0053] S53033: If the sample accuracy score is greater than or equal to the preset score threshold, obtain the instantaneous reward for the model's selected action based on the current task queue of the sample recognition model and the sample accuracy score.

[0054] For example, the instantaneous reward can be obtained using the following formula: ; in, For instant reward, The number of tasks in the current task queue of the sample recognition model. Adjust the coefficients for the task. Score the accuracy of the sample. The preset scoring threshold is defined as follows.

[0055] In a feasible embodiment, S5306: The step of updating the network parameters of the value network and the policy network according to the training samples stored in the experience pool to obtain the agent model includes: S53061: Based on the instantaneous reward and estimated expected value of the training samples, obtain the advantage function, including: The advantage function is obtained using the following formula: ; in, For the aforementioned advantage function, For environmental conditions, Select actions for the model. For the network parameters of the policy network, For the network parameters of the value network, The number of training samples, As a discount factor, The instantaneous reward for the corresponding training sample. This is the estimated expected value.

[0056] S53062: Update the network parameters of the value network and the policy network according to the advantage function, including: The network parameters of the value network are updated using the following formula: ; in, The updated network parameters of the value network, The network parameters of the value network before the update. The learning rate of the value network. The number of training samples, For partial derivatives, This is the advantage function.

[0057] The network parameters of the policy network are updated using the following formula: ; in, The updated network parameters of the policy network, The network parameters of the policy network before the update. The number of training samples, For splitting operators, The output of the policy network, For the aforementioned advantage function, The strength coefficient of the entropy regularization term. Let be the entropy of the policy distribution.

[0058] In this embodiment, by constructing a multi-model collaborative recognition architecture, and considering the diversity and complexity of handwritten content, the system combines the environmental states of multiple lightweight handwritten content recognition models and uses an intelligent agent model to dynamically match the optimal recognition model, which can significantly improve the accuracy and efficiency of handwritten content recognition.

[0059] In summary, the image recognition method for archival information archiving proposed in this application can also be applied to software systems. Specifically, during the image recognition process of the target image of the archive to be processed, the system can first use an OCR module to extract printed text fields (such as title, author, date, document number, project number, Lingnan literature title, etc.), and then use a lightweight handwritten content recognition model to accurately recognize annotations, signatures, and manuscript information. After recognition, keywords are extracted from the recognized content, and the system then automatically categorizes the archives into corresponding digital folders according to custom archiving classification rules, and stores them in a structured manner in the corresponding database.

[0060] For example, when organizing local documents and ancient book manuscripts, the system identifies printed titles, authors, edition dates, edition types, printing information, format, and printing marks. It also identifies handwritten inscriptions, annotations, collation, copies, book collection marks, circulation, and handwritten records. Then, according to custom classification rules, the documents are archived according to the rule of "Lingnan Culture - Document Category - Specific Subject - Historical Period - Document Type - Content Theme" to form a structured digital archive of Lingnan culture to support academic research.

[0061] For example, in industry-education integration training, students upload scanned copies of their creative drafts, manuscripts, and teachers' handwritten annotations and signatures. The system automatically recognizes the titles and content of the printed text, while also detecting the manuscripts, handwritten comments, signatures, and dates, marking them as "correction feedback." After recognition, the system automatically archives and generates structured data according to the classification rules pre-set by the teachers in the system backend, and synchronously links it to the "Course Files - Student Names - Training Reports" database. On the one hand, this creates readily searchable real course files and student profiles, providing students with efficient learning of intelligent file management processes during practical exercises. On the other hand, enterprises obtain a batch of labeled data and model feedback validated in real-world scenarios, forming a closed loop of industry-education integration from "teaching to training to technology optimization," and providing data support for subsequent academic research.

[0062] The above applications effectively solve the problems of low efficiency, poor standardization, insufficient adaptation of cultural archives, and difficulty in supporting industry-education integration training in traditional manual archiving. They transform the technical process into professional digital training content and provide a practical intelligent archiving solution for the research and dissemination of Lingnan culture.

[0063] Please see Figure 3 The second embodiment of this application provides an image recognition device 100 for archiving information, comprising: Clustering analysis module 101 is used to perform clustering analysis on image samples of multiple handwritten contents to obtain multiple classification clusters; The accuracy score acquisition module 102 is used to input image samples of multiple handwritten contents into multiple lightweight handwritten content recognition models, obtain handwritten content recognition result samples output by each lightweight handwritten content recognition model, and the accuracy score of each lightweight handwritten content recognition model for each classification cluster. Image recognition module 103 is used to perform image recognition on the target image of the file to be processed, and obtain the printed text content result of the target image and the area image of multiple handwritten contents; The target classification cluster acquisition module 104 is used to perform cluster analysis on the region images of the plurality of handwritten content to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the plurality of classification clusters; The target recognition model acquisition module 105 is used to select the target recognition model of the region image from the plurality of lightweight handwritten content recognition models according to the accuracy score corresponding to the target classification cluster. The handwriting content recognition module 106 is used to input the images of each region into the corresponding target recognition model to perform handwriting content recognition and obtain the handwriting content recognition results of each region image. The electronic archive content acquisition module 107 is used to acquire the electronic archive content of the archive to be processed based on the printed text content result and the handwritten content recognition result.

[0064] It should be noted that the image recognition device 100 for archival information archiving provided in the second embodiment of this application is only illustrated by the above-described division of functional modules when executing the image recognition method for archival information archiving. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the image recognition device 100 for archival information archiving provided in the second embodiment of this application and the image recognition method for archival information archiving in the first embodiment of this application belong to the same concept, and its implementation process is detailed in the method embodiment, which will not be repeated here.

[0065] The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0066] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0067] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function selected in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function selected in one or more boxes.

[0069] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0070] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0071] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0072] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0073] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An image recognition method for archiving information systems, characterized in that, include: Cluster analysis was performed on image samples containing multiple handwritten text to obtain multiple classification clusters; Image samples of multiple handwritten contents are input into multiple lightweight handwritten content recognition models to obtain handwritten content recognition result samples output by each lightweight handwritten content recognition model, as well as the accuracy score of each lightweight handwritten content recognition model for each classification cluster. Image recognition is performed on the target image of the file to be processed to obtain the printed text content of the target image and the area images of multiple handwritten contents; Cluster analysis is performed on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the multiple classification clusters; Based on the accuracy score corresponding to the target classification cluster, a target recognition model for the region image is selected from the plurality of lightweight handwritten content recognition models. Each of the aforementioned regions is input into the corresponding target recognition model to perform handwritten content recognition, thereby obtaining the handwritten content recognition results for each region image. Based on the printed text content results and the handwritten content recognition results, the electronic archive content of the file to be processed is obtained.

2. The image recognition method for archiving information systems according to claim 1, characterized in that, The step of selecting a target recognition model for the region image from the plurality of lightweight handwritten content recognition models based on the accuracy score corresponding to the target classification cluster includes: The task queue of the lightweight handwritten content recognition model with the highest accuracy score corresponding to the target classification cluster; If the number of queued tasks in the task queue is less than or equal to a preset threshold, the lightweight handwritten content recognition model with the highest accuracy score will be determined as the target recognition model. If the number of queued tasks in the task queue is greater than the preset threshold, an allocation request is sent to the trained agent model to obtain the target recognition model selected by the agent model.

3. The image recognition method for archiving information systems according to claim 2, characterized in that, The agent model includes a policy network and a value network; the agent model is trained through the following steps: Based on multiple image samples of handwritten content, generate multiple sample allocation requests and send them continuously and randomly to the initialized agent model; Upon receiving a sample allocation request, the policy network performs a model selection action based on the current environment state, selecting a sample recognition model from among the various lightweight handwritten content recognition models and updating the environment state; wherein, the current environment state includes the current task queue of each lightweight handwritten content recognition model and the accuracy score corresponding to each classification cluster; Based on the environmental conditions, obtain the instantaneous reward for the action selected by the model; The estimated expected value of the selected action is obtained through the value network; The environmental state, model selection action, instantaneous reward, and estimated expected value corresponding to each allocation request sample are stored as training samples in the experience pool. The network parameters of the value network and the policy network are updated based on the training samples stored in the experience pool to obtain the agent model.

4. The image recognition method for archiving information systems according to claim 3, characterized in that, The step of obtaining the instantaneous reward of the model's selected action based on the environmental state includes: From the current environment state, obtain the sample accuracy score of the sample recognition model for the classification cluster to which the image sample corresponding to the received sample allocation request belongs; If the sample accuracy score is less than a preset score threshold, the instantaneous reward is determined to be a negative reward. If the sample accuracy score is greater than or equal to the preset score threshold, the instantaneous reward for the model's selected action is obtained based on the current task queue of the sample recognition model and the sample accuracy score.

5. The image recognition method for archiving information systems according to claim 4, characterized in that, The step of obtaining the instantaneous reward for the action selected by the model based on the current task queue of the sample recognition model and the sample accuracy score includes: The instantaneous reward is obtained using the following formula: ; in, For instant reward, The number of tasks in the current task queue of the sample recognition model. Adjust the coefficients for the task. Score the accuracy of the sample. The preset scoring threshold is defined as follows.

6. The image recognition method for archiving information systems according to claim 3, characterized in that, The step of updating the network parameters of the value network and the policy network based on the training samples stored in the experience pool to obtain the agent model includes: The advantage function is obtained based on the instantaneous reward and estimated expected value of the training samples; The network parameters of the value network and the policy network are updated according to the advantage function.

7. The image recognition method for archiving information systems according to claim 6, characterized in that, The step of obtaining the advantage function based on the instantaneous reward and estimated expected value of the training samples includes: The advantage function is obtained using the following formula: ; in, For the aforementioned advantage function, For environmental conditions, Select actions for the model. For the network parameters of the policy network, For the network parameters of the value network, The number of training samples, As a discount factor, The instantaneous reward for the corresponding training sample. This is the estimated expected value.

8. The image recognition method for archiving information systems according to claim 6, characterized in that, The step of updating the network parameters of the value network and the policy network according to the advantage function includes: The network parameters of the value network are updated using the following formula: ; in, The updated network parameters of the value network, The network parameters of the value network before the update. The learning rate of the value network. The number of training samples, For partial derivatives, This is the advantage function.

9. The image recognition method for archiving information systems according to claim 6, characterized in that, The step of updating the network parameters of the value network and the policy network according to the advantage function includes: The network parameters of the policy network are updated using the following formula: ; in, The updated network parameters of the policy network, The network parameters of the policy network before the update. The number of training samples, For splitting operators, The output of the policy network, For the aforementioned advantage function, The strength coefficient of the entropy regularization term. Let be the entropy of the policy distribution.

10. An image recognition device for archiving information systems, characterized in that, include: The clustering analysis module is used to perform clustering analysis on image samples of multiple handwritten contents to obtain multiple classification clusters; The accuracy score acquisition module is used to input image samples of multiple handwritten contents into multiple lightweight handwritten content recognition models, obtain handwritten content recognition result samples output by each lightweight handwritten content recognition model, and the accuracy score of each lightweight handwritten content recognition model for each classification cluster. The image recognition module is used to perform image recognition on the target image of the file to be processed, and obtain the printed text content of the target image and the area image of multiple handwritten contents; The target classification cluster acquisition module is used to perform cluster analysis on the multiple handwritten content region images to determine the target classification cluster to which each region image belongs; wherein, the target classification cluster is one of the multiple classification clusters; The target recognition model acquisition module is used to select the target recognition model of the region image from the plurality of lightweight handwritten content recognition models according to the accuracy score corresponding to the target classification cluster. The handwritten content recognition module is used to input the images of each region into the corresponding target recognition model to perform handwritten content recognition and obtain the handwritten content recognition results of each region image. The electronic archive content acquisition module is used to acquire the electronic archive content of the archive to be processed based on the printed text content result and the handwritten content recognition result.

Images