An information representation capability evaluation-based target image screening method and device, an electronic device, and a storage medium

Abstract

The application discloses a target image screening method and device based on information representation capability evaluation, electronic equipment and storage medium. The method comprises the following steps: acquiring a candidate information carrier image, and determining corresponding defect state representation information; determining spatial weighting information according to the defect state representation information, and fusing image features to obtain fused image features; determining information missing area representation results according to the fused image features, and determining information representation capability evaluation results based on the information missing area representation results; and determining a target image according to the information quality adaptation degree corresponding to a plurality of candidate information carrier images. The application evaluates the information expression capability of the candidate image by comprehensively considering the image defects, information loss and information correlation relationship maintenance, avoids screening the target image only according to the image definition or resolution, and improves the adaptability of the target image to subsequent information processing tasks.

Description

Technical Field

[0001] This application relates to the fields of education and training and image processing technology, and in particular to a method, apparatus, electronic device and storage medium for target image screening based on information representation capability evaluation. Background Technology

[0002] With the development of image acquisition equipment and image processing technology, applications based on image information acquisition and processing are gradually increasing. In application scenarios such as educational resource digitization, test bank construction, and textbook content management, it is often necessary to acquire images of information carriers such as paper textbooks, test papers, and workbooks, and then perform processing such as text recognition, formula parsing, and content retrieval based on the acquired images. To improve the effectiveness of subsequent information processing, existing technologies typically evaluate the quality of multiple candidate images and select images with higher clarity, higher resolution, or better image integrity as the objects for further processing.

[0003] However, in practical applications, there is no perfect correlation between image quality metrics and the information-carrying capacity of an image. Specifically, images with high resolution and good visual effects are not necessarily the most suitable target images for information processing input. For example, during the acquisition of images for textbooks, exam papers, etc., factors such as shooting angle, lighting changes, partial occlusion, dirt, or reflections may cause some candidate images, while possessing high resolution and good overall clarity, to lose key text, mathematical symbols, formula structures, or graphical relationships. Conversely, some images with relatively lower resolution, although having reduced overall visual quality, retain the integrity of key characters, the relationships between symbols, and the information structure, making them more suitable as input for subsequent recognition and content understanding. Existing technologies typically rely on single or combined image quality metrics for screening, making it difficult to accurately determine whether an image can effectively express its information content.

[0004] Furthermore, for information carrier images containing different types of information, such as text, formulas, and graphics, the requirements for image quality factors vary depending on the type of information expression. For example, textual information relies more on the integrity of character edges and the arrangement of characters, formula information relies more on the combination of symbols and the preservation of structural relationships, while graphic information relies more on the continuity of contours and structural integrity. Therefore, selecting target images solely based on traditional image quality indicators such as resolution and sharpness is insufficient to reflect the actual expressive ability of images for different types of information, easily leading to the problem of selecting images with high visual quality but poor information recognition performance. Therefore, how to evaluate the information representation ability of candidate information carrier images by combining image defect status and information expression characteristics, and determine suitable candidate images as target images based on the matching relationship between image quality and information representation ability, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a target image screening method, device, electronic device and storage medium based on information representation capability evaluation, to solve the problem that in the prior art, target images are usually selected based on image quality indicators such as resolution and sharpness, which leads to the inability to accurately determine whether the image can effectively express the information it carries.

[0006] To achieve the above objectives, this application provides a target image screening method based on information representation capability evaluation, comprising: Multiple candidate information carrier images are acquired, and defect state characterization information is determined based on the image defect state corresponding to the candidate information carrier images. Based on the defect state characterization information, the information quality evaluation result corresponding to each candidate information carrier image is determined, and the image to be processed is determined from multiple candidate information carrier images based on the information quality evaluation result; Information content reconstruction processing is performed on the image to be processed to obtain a reconstructed image, and the information preservation evaluation result is determined based on the information preservation relationship between the reconstructed image and the image to be processed. Based on the information preservation evaluation results and the image quality evaluation results corresponding to the reconstructed image, the information quality adaptation degree corresponding to the reconstructed image is determined, and the target processing image is determined based on the information quality adaptation degree. A standard image is generated based on the target image, and the standard image is stored in an image database.

[0007] Optionally, determining the defect state characterization information includes: For different defect influencing factors in the image to be processed, corresponding defect influence characterization information is determined, and multiple defect influence characterization information are correlated and fused to generate defect state characterization information that characterizes the overall defect influence degree of the image to be processed.

[0008] Optionally, when determining the defect state characterization information, the degree of defect influence corresponding to each spatial location in the image to be processed is determined based on the defect influence characterization information corresponding to different spatial locations, so as to form spatially distributed defect state characterization information.

[0009] Optionally, based on the defect state characterization information, a weighted fusion of image features corresponding to the image of the information carrier to be processed is performed, including: Based on the defect state characterization information corresponding to different spatial locations, the degree of influence of defects on the image features at the corresponding spatial locations is determined, and spatial weighted information is generated based on the degree of influence of defects on the image features. The image features corresponding to different spatial locations are weighted and fused according to the spatial weighting information to generate the fused image features.

[0010] Optionally, generating the spatial weighting information includes: Based on the degree of influence of defects on image features corresponding to different spatial locations, the degree of influence of defects on the image features is normalized. The feature weights of the corresponding spatial locations are determined based on the normalization results to generate the spatial weighted information.

[0011] Optionally, based on the fused image features, information integrity analysis is performed on the information expression regions in the information carrier image to be processed, including: Based on the information expression type corresponding to the information expression region, determine the information association features corresponding to the information expression region; Based on the information association features, perform information integrity analysis on the information expression region to determine the information loss features corresponding to the information expression region.

[0012] Optionally, based on the information association characteristics corresponding to the information expression type, information integrity analysis is performed on the information expression region, including: Based on the information expression type, a corresponding information association feature is determined. The information association feature is used to characterize at least one of the positional association relationship, combination association relationship, and structural association relationship between different information units in the information expression region. Based on the information association features, the information loss features corresponding to the information expression region are determined.

[0013] Optionally, determining the representation result of the information-missing region includes: The information loss status of the corresponding information expression region is determined based on the information loss characteristics. Based on the spatial positional relationship of the information expression region in the image of the information carrier to be processed, the information loss state is mapped to the corresponding image region to generate the representation result of the information missing region.

[0014] Optionally, determining the information representation capability evaluation result includes: Based on the characterization results of the missing information regions, the information expression regions affected by the missing information are determined; Based on the information association characteristics corresponding to the information expression regions affected by the missing information, the completeness of the association relationship between different information units is determined; The evaluation result of the information representation capability is determined based on the completeness of the aforementioned relationships.

[0015] Optionally, determining the information representation capability evaluation result includes: Based on the multiple information association features corresponding to the information expression region, determine the completeness of the corresponding multiple association relationships; The integrity levels of multiple relationships are fused to generate a relationship integrity fusion result. The evaluation result of the information representation capability is determined based on the fusion result of the association integrity.

[0016] Optionally, determining the target image includes: Obtain image quality characterization information corresponding to the candidate information carrier image; Based on the image quality characterization information and the information characterization capability evaluation results, the information quality matching relationship corresponding to the candidate information carrier image is determined; The target image is determined based on the information quality matching relationship between multiple candidate information carrier images.

[0017] Optionally, determining the information quality matching relationship includes: Based on the information expression type corresponding to the information expression region in the candidate information carrier image, determine the quality characterization parameter in the image quality characterization information that corresponds to the information expression type; The information quality matching relationship is determined based on the quality characterization parameters and the information characterization capability evaluation results.

[0018] Optionally, determining the quality characterization parameters includes: Based on the information association features corresponding to the information expression type, determine the quality characterization parameters corresponding to the information expression region; The quality characterization parameters are used to characterize at least one of the following in the information expression region: the clarity of information units, the degree of preservation of positional relationships between information units, the degree of preservation of combination relationships, and the degree of preservation of structural relationships.

[0019] Optionally, the target image is determined based on the information quality matching relationship corresponding to multiple candidate information carrier images, including: Based on the image quality characterization information and the information characterization capability evaluation results, the information quality fit degree corresponding to the candidate information carrier image is determined; wherein, the information quality fit degree is used to characterize the matching relationship between the image quality features and information expression capability of the candidate information carrier image; The target image is determined based on the degree of information quality adaptation of multiple candidate information carrier images.

[0020] This application also provides a target image screening device based on information representation capability evaluation, comprising: The defect status analysis module is used to acquire multiple candidate information carrier images and determine defect status characterization information based on the image defect status corresponding to the candidate information carrier images. The information quality evaluation module is used to determine the information quality evaluation result corresponding to each of the candidate information carrier images based on the defect state characterization information, and to determine the image to be processed from multiple candidate information carrier images based on the information quality evaluation result; The information content reconstruction module is used to perform information content reconstruction processing on the image to be processed to obtain a reconstructed image, and to determine the information preservation evaluation result based on the information preservation relationship between the reconstructed image and the image to be processed. The information preservation evaluation module is used to determine the information quality fit degree corresponding to the reconstructed image based on the information preservation evaluation result and the image quality evaluation result corresponding to the reconstructed image, and to determine the target processing image based on the information quality fit degree. A standard image generation unit is used to generate a standard image based on the target processed image and store the standard image in an image database.

[0021] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps of the above method.

[0022] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method.

[0023] The target image screening method, apparatus, electronic device, and storage medium based on information representation capability evaluation provided in this application have at least the following beneficial effects: First, this application acquires defect state characterization information corresponding to the image of the information carrier to be processed, and performs weighted fusion of image features corresponding to different spatial locations based on the defect state characterization information, so that image features at different spatial locations can be processed differently according to the corresponding defect state. Compared with the existing technology that uses overall image quality indicators to uniformly evaluate images, this application can reduce the interference of local defects on the overall information characterization ability evaluation results and improve the accuracy of information expression ability characterization for information carrier images with local defects.

[0024] Second, this application analyzes the integrity of information representation areas and determines information association features based on the information representation types corresponding to those areas. This allows the image evaluation process to go beyond visual quality indicators such as pixel sharpness and resolution, and further consider the positional, combinatorial, and structural relationships between information units. When parts of an image are missing, occluded, or blurred, this application can determine whether the defect disrupts the relationships between information units, thereby more accurately determining the actual preservation of the information carried by the image.

[0025] Third, this application determines the evaluation result of information representation capability based on the representation results of information-deficient regions, transforming the target image selection process from a traditional image quality evaluation to an evaluation process oriented towards information expression capability. Since there is not a complete correspondence between image visual quality and information expression capability, this application can avoid the information loss problem caused by selecting target images solely based on factors such as high resolution and high definition, thereby improving the adaptability of the determined target images to subsequent information recognition, content parsing, and other processing tasks.

[0026] Fourth, this application determines the information quality matching relationship by combining the image quality representation information corresponding to the candidate information carrier image and the information representation capability evaluation results, and further determines the degree of information quality fit, so that the target image selection process can comprehensively consider the relationship between image visual quality and information expression capability. Therefore, even if a candidate image is not advantageous in terms of traditional image quality indicators such as resolution and sharpness, it can still be identified as a more suitable target image for subsequent processing if its key text, formulas, graphic structures, and other information expression relationships remain relatively complete.

[0027] Fifth, this application is applicable to image processing scenarios for information carriers containing multiple forms of information expression. For example, in the digitization of educational materials, for textbook and test paper images containing text, mathematical formulas, graphics, etc., this application can evaluate the information preservation capability of the image based on the information association relationship corresponding to different information expression types, and select the target image that is more suitable for text recognition, formula parsing, and content understanding from multiple candidate acquired images, thereby improving the information reliability in the digitization process. Attached Figure Description

[0028] Figure 1 A schematic flowchart of a standard image generation method based on information preservation evaluation provided in an embodiment of this application; Figure 2 A schematic diagram illustrating an application scenario of the standard image generation method based on information preservation evaluation provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a standard image generation apparatus based on information preservation evaluation provided in an embodiment of this application.

[0029] Figure 4 This is a schematic diagram of the electronic device structure provided in the embodiments of this application; Figure 5 This is a schematic diagram of a computer-readable storage medium structure provided in an embodiment of this application. Detailed Implementation

[0030] The technical solutions in the embodiments of this application are clearly and completely described below with reference to the accompanying drawings. It should be noted that the described embodiments are only for explaining this application and are not intended to limit the scope of protection of this application. Those skilled in the art, based on their understanding of the technical solutions of this application, can make adaptive adjustments or substitutions according to actual application needs, all of which should be included within the scope of protection of this application.

[0031] This application provides a standard image generation method, apparatus, electronic device, and storage medium based on information preservation evaluation. The method can be applied to application scenarios requiring automatic screening, processing, and generation of standard image resources from large amounts of non-standardized image data, such as educational material digitization, intelligent document management, knowledge base construction, image resource organization, and intelligent retrieval systems.

[0032] In practical applications, with the development of mobile terminal devices, online platforms, and image acquisition technologies, an increasing number of information resources are being converted into digital images through user photography, scanning, or uploading. For example, in the education field, students, teachers, or other users can use mobile terminals to photograph textbook pages, exam paper pages, workbook pages, or other learning materials, and upload the images to an educational platform for subsequent processing such as question recognition, automatic grading, knowledge point analysis, and learning resource management. In the field of document management, users can also convert paper information into digital images by photographing paper contracts, technical documents, and instruction manuals, further building a searchable and analyzable information resource database.

[0033] However, unlike standard images obtained through professional scanning equipment or manual processing, user-uploaded images typically exhibit significant uncertainty. Due to differences in the acquisition equipment, shooting environments, and shooting habits used by different users, a large number of candidate images with markedly varying quality may be generated for the same information object. For example, for the same textbook page, different users may upload multiple photos, some of which may show the complete page and clear text content, while others may have issues such as missing page edges, fingers obscuring the image, blurry images, tilted angles, or uneven lighting.

[0034] In application scenarios requiring the construction of standardized image databases, not all user-uploaded images have the same value. For example, an image with high overall resolution may be unsuitable as a standard resource if key areas of the page are obscured, resulting in missing conditions, formula symbols, or graphic structures. Conversely, an image with slight brightness variations but complete text, formulas, and illustrations may be more suitable as foundational data for subsequent recognition and understanding. Therefore, automatically identifying images with high information preservation capabilities from a large number of non-standardized user-generated images and further generating standardized images suitable for database storage and intelligent processing is a crucial problem that needs to be addressed in the current construction of digital resources.

[0035] In existing technologies, one approach is to determine standard images through manual screening. For example, staff members review multiple user-uploaded images and select those with better visual quality based on subjective experience. However, when the number of images to be processed is large, manual screening is not only inefficient but also easily affected by staff experience, judgment criteria, and work status, making it difficult to guarantee consistency between different batches of data. Another approach is to automatically select images based on a single image quality metric. For example, the best image is determined based on metrics such as image resolution, sharpness, brightness, or file size. However, a single visual quality metric cannot accurately reflect the image's suitability for subsequent information processing. For example, high-resolution images may have partial occlusion, leading to missing useful information; high-sharpness images may have incomplete page content due to insufficient shooting range; and images with good overall brightness may have key characters that are unrecognizable due to localized reflections. Therefore, relying solely on image surface quality cannot accurately determine whether an image is suitable as a standard image resource.

[0036] Furthermore, in some application scenarios, the images that users upload as information carriers also contain information added by the user later. For example, on textbook pages or test paper pages, users may write answers, solution processes, highlighting key points, or other annotations. While this information has value for individual users, it is not the target information for constructing standard textbook pages or standard question resources. Therefore, it is necessary to remove user-added content and retain only the original content of the information carrier.

[0037] However, directly removing content from user-uploaded images also presents problems. On one hand, user-added content may overlap with the original content, such as handwritten answers covering printed text or handwritten auxiliary lines covering graphic structures. On the other hand, the original image itself may already have issues such as occlusion, blurring, or missing parts, making it difficult for the processing model to accurately determine which content needs to be deleted and which needs to be retained. For example, on a math textbook page, users might add auxiliary lines next to geometric figures. If the processing model judges based solely on line color, thickness, or position, it might delete both the original geometric boundary and the user-added auxiliary lines, resulting in a final image that fails to convey the meaning of the original problem. In the formula area, users might circle a symbol or add calculations; if the processing is inaccurate, it might mistakenly delete subscripts, operators, or key symbols in the formula.

[0038] Therefore, this application proposes a standard image generation method based on information preservation evaluation. This method does not directly select the visually best image from multiple candidate images. Instead, from an information processing perspective, it first analyzes the inherent defects in the candidate images and filters images suitable for further processing based on the degree of impact of these defects. Then, it performs information content reconstruction processing on the filtered images to remove non-target information and restore potentially affected areas. Furthermore, by comparing the information preservation before and after reconstruction, it determines whether the reconstruction process causes a loss of original effective information. Finally, based on the comprehensive evaluation results, it determines the target standard image and stores it in a database.

[0039] The "information content reconstruction processing" in this application is not limited to a single specific processing method. In a preferred embodiment, handwriting removal can be used as a specific implementation method for information content reconstruction, that is, identifying and removing handwritten content added by the user to the original information carrier image while preserving the original printed text, formulas, illustrations, and page structure information. In other embodiments, the information content reconstruction processing may also include damaged area repair, missing area restoration, local content enhancement, and other processing methods to improve the usability of the information carrier image.

[0040] The method of this application may generally include the following steps: Step S1: Acquire multiple candidate information carrier images, and determine the defect state characterization information based on the defect state corresponding to the candidate information carrier images.

[0041] Step S2: Determine the information quality evaluation results corresponding to multiple candidate information carrier images based on the defect state characterization information, and determine the image to be processed based on the information quality evaluation results.

[0042] Step S3: Perform information content reconstruction processing on the image to be processed to obtain a reconstructed image.

[0043] Step S4: Determine the information preservation evaluation result based on the information preservation relationship between the reconstructed image and the image to be processed.

[0044] Step S5: Determine the target standard image based on the evaluation results and store the target standard image in the database.

[0045] Step S1 is mainly used to address the problem of large differences in the quality of candidate images. By analyzing the defect states in different candidate images, it provides a basis for subsequent image selection. Step S2 is mainly used to determine the images suitable for entering the processing flow based on the defect states, so as to avoid low-quality input affecting the subsequent processing effect. Step S3 is mainly used to reconstruct the non-target information in the image to be processed, so that the image better meets the requirements of standard resources. Step S4 is mainly used to evaluate whether the reconstruction process retains the original valid information. Step S5 is mainly used to determine the final standard image based on the comprehensive evaluation results, so as to realize the automatic construction of standard image resources.

[0046] Through the above processing flow, this application forms a closed-loop processing mechanism from input screening and content reconstruction to result evaluation. Specifically, the initial stage uses defect state analysis to reduce the impact of low-quality input on subsequent processing, while the subsequent stage uses information preservation evaluation to reduce the risk of information loss during content reconstruction, thereby improving the reliability of the final generated standard image.

[0047] S1. Acquire multiple candidate information carrier images, and determine the defect state characterization information based on the defect state corresponding to the candidate information carrier images.

[0048] In one implementation, this application first acquires multiple candidate information carrier images. These candidate information carrier images can be multiple image data collected from the same information object. For example, in a scenario of digitizing educational materials, multiple users can each use a mobile terminal to photograph the same textbook page, test paper page, or workbook page, and upload the photographed images to a server; in a scenario of digitizing documents, multiple users can each photograph the same paper document, technical document, or instruction manual, thereby forming multiple candidate information carrier images.

[0049] Because multiple candidate information carrier images originate from different acquisition processes, there are often quality differences between them. Factors causing these quality differences can include variations in acquisition device performance, shooting distance, shooting angle, ambient lighting, user operation methods, and quality changes during subsequent image compression and transmission. For example, the same textbook page may correspond to multiple user-uploaded photos. One photo may completely cover the entire page with clear text, another may have missing page edges due to insufficient shooting range, and yet another may have areas that are unrecognizable due to finger obstruction or changes in lighting.

[0050] In application scenarios that require building a standard image database, candidate information carrier images cannot be directly used as final data resources. On the one hand, some candidate images may already have missing information, such as characters being missing due to page cropping or content being invisible due to occlusion; on the other hand, although some candidate images are visually clear, they may not meet the needs of subsequent text recognition, formula parsing, or content understanding due to degraded quality in local areas.

[0051] Therefore, in one implementation, defect state analysis can be performed on multiple candidate information carrier images to determine the defect state characterization information corresponding to each candidate information carrier image. The defect state characterization information describes the type of defect, the degree of defect impact, and the spatial distribution of defects in the candidate information carrier images. Specifically, a defect detection model can be used to analyze the candidate information carrier images. The defect detection model can be used to identify image defects that affect the subsequent processing effect of the information carrier images and output the corresponding defect type and defect state information.

[0052] In one implementation, the defect detection model may include a basic feature extraction structure and multiple defect analysis branches. The basic feature extraction structure extracts multi-level image features from the candidate information carrier image; the multiple defect analysis branches perform feature analysis on different types of defects and output corresponding defect detection results. The basic feature extraction structure can employ a convolutional neural network, a visual Transformer network, or other network structures capable of image feature extraction. After inputting the candidate information carrier image, the basic feature extraction structure can extract image features from different scales. For example, low-level features can describe edges, textures, brightness variations, and local structural information in the image; high-level features can describe page layout, text region relationships, and overall spatial structure information. By combining image features at different levels, the defect detection model can not only focus on pixel changes on the image surface but also understand the impact of defects on information expression. For example, a blurry defect of the same degree may have a smaller impact if it occurs in a blank area of ​​the page; however, if it occurs in a text area or formula area, it may lead to subsequent recognition failure. Therefore, it is necessary to combine the semantic information of the image region to determine the degree of defect impact.

[0053] It should be noted that the defect detection model in this application is only an execution carrier for implementing the defect state analysis process of the candidate information carrier image. It is used to determine the corresponding defect state information based on the input candidate information carrier image. The technical solution of this application is not intended to protect any specific model structure, network architecture or model training method.

[0054] Specifically, this application focuses on the process of performing defect state analysis on candidate information carrier images. This involves analyzing the candidate information carrier images to determine the defect types, defect confidence levels, and spatial distribution information that affect subsequent information content processing, and generating defect state characterization information based on the analysis results. This defect state characterization information reflects the reliability of information in different spatial regions of the candidate information carrier images and is further used in subsequent image selection, information content reconstruction, and information retention evaluation processes.

[0055] In one implementation, the defect detection model can be implemented using a deep learning-based image analysis model, such as a processing structure based on a convolutional neural network, a visual Transformer network, or other visual feature analysis models. The defect detection model can extract features from candidate information carrier images to obtain texture features, edge features, spatial structure features, and semantic association features, and determine the defect state corresponding to the candidate information carrier image based on these features.

[0056] In other embodiments, the defect detection model can also be implemented using traditional image processing algorithms, rule-based algorithms, or other data analysis methods. For example, abnormal regions in candidate information carrier images can be determined through image segmentation, edge detection, texture analysis, brightness change analysis, etc., and the defect status can be determined based on the location, range, and corresponding content type of the abnormal regions. Therefore, this application does not limit the specific type, network structure, parameter settings, or training method of the defect detection model.

[0057] Furthermore, in embodiments employing machine learning or deep learning models for defect state analysis, the model can be trained using pre-constructed sample data to learn the correlation between candidate information carrier images and corresponding defect states. However, it should be noted that the aforementioned model training process is not the focus of this application. This application does not limit the construction method of training samples, the design method of the model loss function, or the optimization method of model parameters. Instead, it focuses on using image analysis to determine the defect state corresponding to the candidate information carrier image and generating defect state representation information that reflects the reliability of the image based on the defect state. In other words, regardless of the model or algorithm used for defect detection, as long as the corresponding defect state can be determined from the candidate information carrier image, and further defect state representation information used to describe the impact of the defect is obtained, it falls within the scope of this application's technical solution.

[0058] Unlike existing technologies that only perform quality scoring or target recognition on images, the defect state analysis process in this application not only focuses on whether there are anomalies in the candidate information carrier image, but also further analyzes the spatial location of the anomalies and their impact on subsequent information processing. For example, for candidate information carrier images with occlusion, this application not only determines the existence of occlusion defects, but also further determines whether the occluded area is located in a text area, formula area, or graphic area, and determines the degree of impact of the corresponding defects based on the information value of different areas. As another example, for candidate information carrier images with blurring defects, this application can further determine whether the blurring area affects the effective expression of information, thereby providing a basis for subsequent image selection and content reconstruction.

[0059] Furthermore, this application does not limit the defect status analysis process to be completed by a single defect detection model. In one embodiment, corresponding analysis models can be set for different types of defects, and multiple analysis results can be fused to generate comprehensive defect status characterization information; in another embodiment, a single comprehensive analysis model can be used to simultaneously output analysis results corresponding to multiple defect types. The different implementation methods described above do not affect the implementation of the technical solution of this application.

[0060] In this way, the present application uses the defect detection model as a technical tool to realize defect state analysis. By performing defect state analysis on the candidate information carrier image, defect state characterization information that can characterize the reliability of the image is obtained, and the defect state characterization information is used to guide the subsequent standard image generation process, thereby reducing the impact of input image defects on the final generation result.

[0061] In one embodiment, the defect type may include at least one of the following: truncated page defect, image occlusion defect, image blurring defect, text distortion defect, angle rotation defect, and uneven lighting defect.

[0062] Among these, truncated page defects describe situations where the candidate information carrier image does not completely cover the original information carrier area. For example, when a user photographs a paper page, insufficient shooting range, page edges exceeding the camera's range, or incomplete subsequent cropping may result in missing areas of the page. This type of defect differs from ordinary image quality degradation; it directly causes some original information to be absent from the candidate image. For instance, if part of the text in a question is cropped, subsequent processing cannot recover the missing characters from the image; missing symbols in a formula may render the entire formula unreadable; missing edges in illustrations may affect the overall understanding of the graphic. Therefore, in the defect status analysis process, it is crucial to identify the completeness of the page boundaries.

[0063] Image occlusion defects describe situations where a portion of an image containing candidate information is covered by other objects. Examples include a user's finger obscuring a page area during photography, a pen or ruler covering text, or handwriting covering the original printed area. Unlike truncated pages, the original information in the occluded area is not completely absent but rather covered by other content, so in some cases it may still be possible to recover it using contextual information. However, when the occluded area is located in the position of key text, formula symbols, or graphic structures, it will severely affect the subsequent information extraction process. Therefore, when identifying occlusion defects, it is necessary not only to determine whether occlusion exists but also to further analyze the location of the occluded area and the corresponding degree of impact.

[0064] Image blurring defects describe a decrease in the ability to express detailed information in an image that serves as a candidate information carrier. These defects can be caused by equipment movement during shooting, autofocus failure, inappropriate shooting distance, or image compression. For example, blurred edges of text strokes may make it impossible to distinguish similar characters; blurred superscripts and subscripts in formulas may alter the meaning of the formula; and blurred fine lines in illustrations may make the graphic structure unrecoverable. Therefore, image blurring defects require a comprehensive assessment considering both the degree of blur and the area where it occurs.

[0065] Text distortion defects describe situations where the structure or arrangement of characters in a text area changes due to perspective variations, page curvature, or changes in the shooting angle. For example, when a user does not photograph a paper perpendicularly, different areas of the page may be stretched or compressed to varying degrees, altering the row and column relationships of the text. While these defects typically do not cause direct pixel loss, they disrupt the text's structure, affecting text detection, character recognition, and layout analysis.

[0066] Angle rotation defects describe situations where the overall orientation or the orientation of a local area of ​​a candidate information carrier image deviates from a preset orientation. For example, if a user's phone is tilted when taking a picture of a page, causing the entire page to rotate; or if a user places paper at an abnormal orientation, the text area cannot be read in the normal direction. These defects can usually be addressed through image correction, but excessive rotation angles can still affect subsequent region localization and content recognition.

[0067] Uneven illumination defects are used to describe situations where the brightness distribution is inconsistent across different areas of an image containing candidate information. For example, uneven light sources in the shooting environment can cause one side of the page to be brighter than the other, resulting in shadows on the other side; or reflections from the paper surface can cause overexposure in some areas. These defects may not directly change the page structure, but they can reduce the recognizability of information in local areas, such as causing character strokes to blend into the background or making graphic boundaries indistinct.

[0068] Through the above defect analysis, defect detection results corresponding to the candidate information carrier image can be obtained. In one embodiment, the defect detection results may include defect type information and corresponding defect confidence information. The defect confidence information is used to indicate the probability that the candidate information carrier image contains a certain type of defect.

[0069] For example, the defect detection model can output: confidence scores for truncated page defects; occlusion defects; blurring defects; text distortion defects; rotation defects; and uneven lighting defects. In this way, different types of defects in candidate images can be converted into calculable evaluation parameters.

[0070] Furthermore, since some defects exhibit distinct spatial distribution characteristics, the overall defect confidence level alone cannot accurately describe the extent of their impact. Therefore, in a preferred embodiment, a defect probability matrix corresponding to the candidate information carrier image can be further determined. This defect probability matrix represents the probability of defect occurrence or the degree of defect impact at different spatial locations within the candidate information carrier image. Specifically, the candidate information carrier image can be divided into multiple spatial regions, and a corresponding defect probability value can be determined for each spatial region.

[0071] For example, for occlusion defects, an occlusion probability matrix can be generated:

[0072] in, This represents the probability that the spatial region in the m-th row and n-th column of the candidate information carrier image is affected by occlusion.

[0073] Similarly, fuzzy probability matrices, illumination anomaly probability matrices, and probability matrices corresponding to other defects can be generated separately.

[0074] The defect probability matrix allows for a transformation from overall defect evaluation to spatial region defect evaluation. For example, two candidate images may both show occlusion defects, but the occluded area in the first image is located in a blank area of ​​the page, while the occluded area in the second image covers the text area of ​​the title. If only the occlusion confidence level is used for judgment, the two may yield similar evaluation results; however, by using the occlusion probability matrix, the information value corresponding to the occluded area can be further determined, resulting in a more accurate description of the defect status.

[0075] Furthermore, in one embodiment, defect state representation information corresponding to candidate information carrier images can be generated based on defect confidence information and defect probability matrices corresponding to multiple defect types.

[0076] The defect status characterization information may include: Global defect characterization information is used to describe the degree to which the overall image of the candidate information carrier is affected by defects; Local defect distribution information is used to describe the degree of impact of defects in different spatial regions; Defect type association information is used to describe the overall impact relationship between different defect types.

[0077] For example, one image may be clear overall but have local occlusions; another image may not have occlusions but is slightly blurry overall. By using defect state characterization information, the different defect characteristics of both can be described simultaneously, providing a more comprehensive data foundation for subsequent quality assessment.

[0078] In one implementation, the defect state characterization information can be further used to determine the information quality evaluation result corresponding to the candidate information carrier image. Specifically, based on the degree of influence of different defect types on the information preservation capability, corresponding influence weights can be set for different defect types, and the comprehensive evaluation result can be determined by combining the defect confidence level.

[0079] For example, the defect evaluation value can be determined in the following way:

[0080] in, This represents the defect evaluation value corresponding to the candidate information carrier image; This represents the influence weight corresponding to the i-th type of defect; This represents the defect confidence level corresponding to the i-th type of defect.

[0081] The weighting parameter reflects the importance of different defect types to subsequent processing. For example, since truncated page defects may lead to the direct loss of original information, a higher weight can be assigned; since minor uneven lighting defects can be improved through subsequent enhancement methods, a relatively lower weight can be assigned. It should be noted that the above evaluation formula is only one implementation method. In other implementation methods, the evaluation results corresponding to different spatial locations can be further adjusted by combining the defect probability matrix, the importance of regional information, and the requirements of subsequent processing tasks. Through the above steps, defect state characterization information corresponding to multiple candidate information carrier images can be obtained, providing a basis for subsequently determining the images suitable for information content reconstruction processing.

[0082] S2. Determine the information quality evaluation result based on the defect status characterization information, and determine the image to be processed based on the information quality evaluation result.

[0083] In one embodiment, after obtaining defect state characterization information corresponding to multiple candidate information carrier images, the information quality evaluation result corresponding to each candidate information carrier image can be further determined based on the defect state characterization information, and the image to be processed can be determined from the multiple candidate information carrier images based on the information quality evaluation result.

[0084] It should be noted that step S1 is mainly used to identify the defect state existing in the candidate information carrier image, that is, to answer "what problems exist in the candidate image"; while step S2 is mainly used to determine whether the candidate information carrier image is suitable as input for subsequent information content reconstruction processing based on the above defect state, that is, to answer "which candidate images are worth entering the next processing stage".

[0085] In practical applications, although different candidate information carrier images may correspond to the same information object, their adaptability to subsequent processing varies significantly due to different acquisition conditions. For example, the first candidate image may be generally clear, but there may be local occlusion on the page; the second candidate image may not have obvious occlusion, but the text area may be slightly blurred due to the shooting distance; the third candidate image may have a complete page, but there may be significant perspective distortion. If all candidate information carrier images are directly input into the subsequent information content reconstruction model, it will increase the consumption of computational resources; on the other hand, low-quality candidate images may affect the accuracy of the reconstruction model's judgment. For example, for images with occlusion defects, the model may not be able to determine whether the occluded area originally belonged to handwritten content or printed content; for images with severe blur defects, the model may not be able to accurately distinguish character edges and noise areas, resulting in the wrong deletion of valid information. Therefore, it is necessary to select suitable images for processing based on the information quality evaluation results of the candidate images before performing information content reconstruction processing.

[0086] In one implementation, the information quality evaluation result corresponding to the candidate information carrier image can be determined based on the defect status characterization information. The information quality evaluation result comprehensively reflects the information preservation capability and applicability of the candidate information carrier image for subsequent processing. Specifically, the information quality of a candidate information carrier image is not entirely equivalent to image clarity in the traditional sense. For example, although an image may have a high resolution, if key text areas are obscured, its information utilization value may be lower than that of an image with a slightly lower resolution but a complete page. Therefore, the information quality evaluation in this application not only considers the visual quality of the image but also comprehensively judges it based on the defect type, the degree of defect impact, and the spatial distribution of defects.

[0087] In one implementation, the corresponding information quality evaluation value can be determined based on the defect state characterization information. Specifically, the defect impact parameters can first be determined based on the degree of impact corresponding to different defect types, and then the information quality evaluation result corresponding to the candidate information carrier image can be calculated by combining the defect confidence information and the defect spatial distribution information.

[0088] For example, for a candidate information carrier image, its information quality evaluation result can be determined according to the following method:

[0089] in, This represents the information quality evaluation result corresponding to the i-th candidate information carrier image; This represents the defect impact value corresponding to the k-th type of defect; This represents the impact weight corresponding to the k-th type of defect; N represents the number of defect types.

[0090] The impact value of a defect can be determined based on its confidence level and spatial distribution. For example, for an occlusion defect, the impact value can be determined based on the occlusion confidence level and the proportion of the occluded area to the effective content area of ​​the image. For a blurry defect, the impact value can be determined based on the degree of blur and whether the blurry area is located in text, formula, or illustration areas. This approach avoids evaluating solely based on the presence of a defect. For instance, both candidate images may show occlusion defects, but the occluded area in the first image is located in the blank area at the edge of the page, having a smaller impact on the effective information; while the occluded area in the second image is located in the main text area, having a larger impact on subsequent recognition and understanding. Simply judging that both images have occlusion would not reflect the actual difference in impact; however, combining spatial distribution information yields an information quality evaluation result that better meets the needs of practical applications.

[0091] In a preferred embodiment, an information region importance ranking can be further introduced to adjust the impact of defects at different spatial locations. Specifically, different regions in an information carrier image have different information values. For example, the title, question stem, formula, and illustration regions on a textbook page typically contain high-value information, while the blank or decorative areas at the page edges have relatively low information value. Therefore, even if two regions are affected by defects of the same degree, their impact on the final standard image generation process may differ. Based on the above reasons, information expression regions in candidate information carrier images can be pre-determined, and region weight maps can be generated according to the importance of different regions.

[0092] The region weight map can be represented as: .in, It indicates the importance of information in the region corresponding to the pixel spatial location (x,y).

[0093] Furthermore, the impact of spatial defects can be determined by combining the defect probability matrix and the region weight map:

[0094] in, This indicates the degree of defect impact corresponding to the pixel spatial location (x, y); This represents the defect probability matrix.

[0095] By employing the methods described above, the evaluation process can focus more on the areas that truly affect the ability to retain information. For example, occlusion defects covering the text area of ​​the question stem can yield a higher defect impact value, while occlusion defects covering blank areas of the page can yield a lower defect impact value.

[0096] In one implementation, the candidate information carrier images can be sorted according to their information quality evaluation results, and the images to be processed can be determined based on the sorting results. Specifically, the candidate information carrier images can be sorted from high to low according to their information quality evaluation results, and the candidate images with the highest sorting results can be selected as the images to be processed. For example, when multiple users upload ten candidate images for the same page, the information quality evaluation result corresponding to each candidate image can be calculated first based on the defect status characterization information, and then the images with the highest evaluation results can be selected to enter the subsequent information content reconstruction process.

[0097] In a preferred embodiment, it is not limited to selecting only one image to be processed, but multiple images can be selected. This is because, in complex application scenarios, different candidate images may have complementary information relationships. For example: First candidate image: The page is clear overall and the text area is complete, but the illustration on the right is slightly obscured.

[0098] Second candidate image: The page is slightly blurry, but the illustration is complete.

[0099] Third candidate image: The text area is slightly distorted, but the page edges are intact.

[0100] If only one image is selected as the processing object, the effective information contained in the other candidate images may not be fully utilized. Therefore, multiple high-quality candidate images can be selected to perform information content reconstruction separately, and the final standard image can be selected in subsequent steps based on the information in the reconstruction results and the evaluation results.

[0101] In one implementation, a threshold filtering method can be used to determine the image to be processed.

[0102] Specifically, an information quality evaluation threshold can be preset. When the information quality evaluation result of a candidate information carrier image meets the above conditions, it can be identified as an image to be processed; otherwise, the candidate image can be considered unsuitable for subsequent processing. The advantage of using the threshold screening method is that it can avoid the impact of changes in the number of candidate images on the processing effect of this application. For example, when a user uploads a large number of low-quality images, this application will not introduce poor-quality data by forcibly selecting a fixed number of images; when a user uploads a small number of high-quality images, all images that meet the requirements can be retained.

[0103] In another implementation, a sorting and thresholding method can be used to determine the images to be processed. For example, candidate images with information quality evaluation results below a threshold can be filtered out first, and then a preset number of images with the highest evaluation results can be selected from the remaining candidate images. This method balances quality constraints and quantity control, making it suitable for application scenarios that require control over computational resource consumption.

[0104] In one implementation, the number of images to be processed can be dynamically adjusted based on the subsequent information content reconstruction processing capability. For example, when subsequent processing resources are sufficient, more candidate images can be selected as images to be processed to obtain more reconstruction results and improve the success rate of generating the final standard image; when computing resources are limited, a small number of candidate images with the highest evaluation results can be selected to reduce processing costs. Furthermore, after determining the images to be processed, a correspondence between candidate information carrier images and images to be processed can be established, and the corresponding defect status characterization information can be input into the subsequent processing flow. In this way, during the subsequent information content reconstruction process, this application can not only obtain the image to be processed itself, but also obtain the defect status information corresponding to the image. For example, if a certain region has been detected with a high occlusion probability in step S1, the subsequent reconstruction process can adopt a more cautious processing strategy for that region; if a certain region does not have obvious defects, the original content of that region can be preserved first. Through the above method, this application does not simply select input data based on image clarity or resolution, but combines defect status, spatial distribution, and information value to determine suitable images to be processed, reducing the impact of low-quality data on the subsequent content reconstruction process from the input stage, and providing a reliable foundation for generating high-quality standard images.

[0105] It should be noted that the above information quality evaluation method is only one implementation method. In other implementation methods, machine learning models, rule-based evaluation models, or other comprehensive evaluation methods can be used to determine the information quality evaluation results corresponding to the candidate information carrier images, depending on the specific application requirements.

[0106] For example, an information quality assessment model can be constructed, taking defect state representation information as input and outputting a corresponding information quality score. In this implementation, historical image data and human evaluation results can be used to train the model, enabling it to learn the correlation between different defect states and image usability.

[0107] It should be emphasized that the focus of this application is not on the specific scoring formula or information quality evaluation model used, but on using the defect status information corresponding to the candidate image to determine its applicability for processing, and on selecting the images to be processed for subsequent information content reconstruction based on the evaluation results.

[0108] S3. Perform information content reconstruction processing on the image to be processed to obtain a reconstructed image.

[0109] In one implementation, after determining the image to be processed based on the information quality evaluation result corresponding to the candidate information carrier image, the image to be processed can be further subjected to information content reconstruction processing to obtain the corresponding reconstructed image.

[0110] It should be noted that the information content reconstruction processing in this application is not a simple image deletion or image enhancement operation, but rather an analysis of the relationship between different types of content in the information carrier image. Based on the identification of non-target information, the corresponding areas are subjected to removal, restoration, and structure preservation processing to obtain reconstruction results that better meet the requirements of standard images.

[0111] In practical applications, images to be processed often contain multiple types of information simultaneously. For example, textbook pages, exam paper pages, or technical document pages typically include printed text, formulas, tables, illustrations, page backgrounds, and information added by the user later. Among these, printed text, formulas, tables, and illustrations are usually inherent content of the information carrier itself, while handwritten answers, annotations, and markings added by the user later are usually non-target information.

[0112] Therefore, when constructing standard image resources, it is necessary to retain the original content of the information carrier and remove information subsequently added by the user. For example, in the process of constructing an educational resource database, it is necessary to retain the question text, formulas, graphics, and page structure, while removing students' written answers, solutions, and annotations. Similarly, in the process of digitizing technical documents, it is necessary to retain the original explanatory text, structural diagrams, and parameter tables, while removing handwritten marks generated by users during reading.

[0113] However, the process of reconstructing information content presents significant technical challenges because different elements within the image to be processed may have spatial overlap. User-added information may not necessarily be located in blank areas but could potentially cover the original content. For example, a user might directly circle key points on the title text, add calculations to formulas, or draw auxiliary lines in a graph. In such cases, if the area to be deleted is determined solely based on color, line thickness, or pixel changes, it's easy to accidentally delete valuable original information along with it.

[0114] Therefore, in one implementation, the image to be processed can be analyzed using content understanding to determine the information categories corresponding to different regions in the image, and different processing strategies can be adopted according to different information categories. Specifically, an information content reconstruction model can be used to perform analysis and processing on the image to be processed. The information content reconstruction model may include an image feature extraction structure, a content region recognition structure, a non-target information detection structure, and a content restoration structure. Among them, the image feature extraction structure is used to obtain multi-level visual features in the image to be processed; the content region recognition structure is used to determine the information types corresponding to different regions in the image; the non-target information detection structure is used to determine the information regions that need to be removed; and the content restoration structure is used to generate reconstruction results based on the information surrounding the region to be processed.

[0115] In one implementation, the image feature extraction structure can be implemented using a deep learning-based visual feature extraction structure. For example, convolutional neural networks, visual Transformer networks, or other model structures capable of extracting spatial and semantic features of images can be used.

[0116] Low-level visual features can be used to describe pixel variations, edge information, texture information, stroke direction, and local color distribution; high-level semantic features can be used to describe character arrangement relationships, page layout relationships, graphic structure relationships, and regional semantic information. By combining features at different levels, the information content reconstruction model can not only identify local visual differences but also understand the relationships between different contents.

[0117] For example, printed text and handwritten text may have similar colors and stroke widths, but they differ in structure. Printed text typically has a regular font structure, stable character spacing, and a fixed arrangement, while handwritten text usually exhibits variations in character size, stroke direction, and irregular arrangement. By combining local visual features with overall structural features, the accuracy of distinguishing between the two can be improved.

[0118] In one implementation, the image to be processed can be divided into content regions to determine the information categories corresponding to different regions. Specifically, the image to be processed can be divided into multiple spatial regions, and each spatial region can be classified. The information categories may include: (1) original printed text region; (2) formula and symbol region; (3) graphic or illustration region; (4) table structure region; (5) user-added information region; and (6) background region. Among them, the original printed text region, formula region, graphic region, and table region are information regions that need to be protected; the user-added information region is the region to be processed; and the background region is used for subsequent content recovery. By using the above region classification method, the problem of determining the deletion region based solely on pixel differences in traditional image processing methods can be avoided.

[0119] For example, on a math textbook page, a user might write their answer next to a problem using a black pen. Since printed text and handwritten text are similar in color, judging solely by color can easily lead to confusion. However, a region-based classification structure allows for further analysis of text arrangement, character structure, and contextual relationships, thus determining whether a region is more likely to belong to printed text or user-added content.

[0120] In a preferred embodiment, the information content reconstruction process can specifically be manifested as handwriting removal processing. The handwriting removal processing is used to identify user handwritten content in the image to be processed, and to perform removal and restoration processing on the user handwritten content to generate a handwriting-removed image. It should be noted that handwriting removal processing is only one specific implementation of information content reconstruction, and this application does not limit information content reconstruction to the removal of handwritten information. In other embodiments, information content reconstruction may also include: repair of damaged areas, completion of missing areas, local content enhancement, background texture restoration, page structure correction, etc., and this application does not limit these aspects. Any method that optimizes and reconstructs the image content while maintaining the original information expression capability can be used as the information content reconstruction processing method of this application.

[0121] The following description uses handwriting removal as a preferred implementation method.

[0122] In one implementation, handwritten regions in the image to be processed can be detected first, and corresponding handwritten region representation information can be generated. Specifically, handwritten region detection is not simply about judging all irregular strokes in the image, but rather about combining image context information to determine whether a region belongs to content added by the user later. For example, on a test paper page, a student might tick a multiple-choice question. Although the tick symbol is simple in form, its position and semantics differ from the original printed content; on a math problem page, a student might write calculation processes next to formulas, which may include numbers, symbols, or even content similar to the original formula; in a graphic area, a student might draw auxiliary lines, which may have similar visual features to the original graphic. Therefore, handwritten region detection needs to comprehensively consider factors such as handwriting shape, spatial position, relationship with the original content, page structure, and contextual semantics.

[0123] In one implementation, a handwritten region probability corresponding to each spatial location can be output by a handwritten region detection branch, and a handwritten region mask can be generated based on the handwritten region probability. The handwritten region mask can be represented as:

[0124] in, This represents the probability that the pixel spatial location (x, y) in the image to be processed belongs to the handwritten region; This indicates a preset judgment threshold. When a certain region corresponds to a higher handwriting probability, it can be considered that the region is more likely to belong to information that needs to be removed; when a certain region corresponds to a lower handwriting probability, it is considered that the region is more likely to belong to the original content to be retained. When =1, it indicates that the corresponding position belongs to the area to be processed; when When =0, it means that the original content is preserved at the corresponding position.

[0125] It should be noted that the mask generation method described above is only one implementation method. In other implementation methods, continuous weights can be used to represent the influence of the handwritten region instead of a binary mask. For example, the handwritten probability can be directly used as the processing weight, allowing the model to perform progressive recovery based on the probability corresponding to different regions.

[0126] In one implementation, after obtaining the handwritten region mask, the mask can be further optimized by incorporating original information protection constraints to reduce the risk of accidental deletion due to spatial overlap between the handwritten content and the original content. Specifically, although the handwritten region detection structure can determine candidate handwritten regions based on visual features, relying solely on handwriting probability may still lead to incorrect judgments in complex page scenarios. For example, printed fonts in textbooks may have discontinuous edges due to blurry images, visually resembling handwriting; symbols in formulas may be mistaken for irregular writing due to their unique font structure; and the line structure in graphics may have a high similarity to user-added auxiliary lines.

[0127] Therefore, in a preferred embodiment, an original content protection mechanism can be further set up to impose secondary constraints on the candidate handwritten area. This original content protection mechanism is used to determine whether the candidate deletion area might contain information content from the original information carrier itself, and adjusts the final processing area based on the determination result.

[0128] Specifically, different protection rules can be determined based on the content type of the original information carrier. For example, for printed text areas, the focus can be on analyzing character continuity, character arrangement patterns, and text context relationships; for formula areas, the focus can be on analyzing symbol combination relationships, subscript and superscript relationships, and mathematical structure relationships; for graphic areas, the focus can be on analyzing line connection relationships, spatial topological relationships, and overall structural integrity. Through these methods, the information content reconstruction process can not only focus on "which content resembles handwriting" but also on "which content cannot be deleted."

[0129] In one implementation, text protection constraints can be established for printed text regions. Specifically, printed text typically has structural characteristics that distinguish it from handwritten content. For example, printed text usually has a stable font shape, regular character spacing, a relatively fixed arrangement direction, and continuous text context. Therefore, when a candidate processing region meets the above-mentioned characteristics of printed text, even if the region has a certain degree of stroke variation, the likelihood of it being deleted can be reduced.

[0130] For example, on an exam paper page, users might circle or mark the text above the questions. If only the dark pixel area is processed, a portion of the question text might be mistaken for handwritten content. However, by analyzing the character arrangement, it can be found that this area still belongs to a continuous text structure, thus increasing its protection weight.

[0131] Furthermore, for images containing formulas, a formula structure protection mechanism can be implemented. In practical applications, formulas typically differ from ordinary text; their information expression relies not only on individual characters but also on the spatial relationships between them. Examples include superscript and subscript relationships, fractional structures, square root structures, operator connections, and symbol arrangement. If some symbols in a formula are deleted during the de-handwriting process, even if the remaining characters still exist, the formula's semantics may change. For instance, a user might write calculations next to a mathematical formula, where some numbers and symbols might overlap with the original formula area. If the model fails to recognize the formula structure, it might misinterpret the numbers or operators in the original formula as handwritten content.

[0132] Therefore, in one implementation, a formula protection region can be established using the formula region detection results. When a region is detected as belonging to a formula region, the content retention priority of that region can be increased, and the decision to perform a deletion operation can be made based on the formula's structural relationships. For example, information such as symbol connections, baseline positions, subscript and superscript positions, and mathematical structure templates within the formula region can be analyzed. If a candidate deletion region would disrupt the aforementioned structural relationships, the corresponding deletion probability can be reduced. Through this formula protection mechanism, the accuracy of reconstruction in scenarios involving complex mathematical materials and technical documents can be improved.

[0133] Furthermore, for information carrier images containing illustrations or graphic information, a graphic structure protection mechanism can be set up. In practical applications, illustrations usually express information through spatial structure, and their importance depends not only on a single pixel or line segment, but also on the overall structural relationships. For example, geometric figures rely on boundary lines and connections; flowcharts rely on nodes and connections; experimental apparatus diagrams rely on the spatial relationships between components; and technical structure diagrams rely on the layout relationships of components. Therefore, simply deleting suspected handwritten lines may destroy the original graphic expression. For example, in a mathematical geometry problem, a student might add auxiliary lines next to the original figure. If the model only judges whether the line belongs to the new content, it may delete both the auxiliary line and the original graphic boundary; however, if the graphic structure is further analyzed, it can be determined which lines belong to the main body of the original figure and which lines belong to the subsequently added content.

[0134] In one implementation, a graphic structure analysis module can determine the structural features corresponding to the graphic region and adjust the handwritten region processing results based on these features. For example, edge connectivity, line segment direction, region closure, and graphic topological relationships can be extracted. When a candidate deletion region is located at a critical structural position, the deletion probability can be reduced. This approach avoids damaging the original illustration during the information content reconstruction process.

[0135] In one implementation, after determining the final area to be processed, content removal and area restoration can be performed. Specifically, the target area requiring information reconstruction can be determined based on the optimized handwritten area mask, and non-target information within the target area can be removed. For example, for handwritten content located in a blank background area, the handwriting can be directly removed, and the page background can be restored based on the surrounding background texture; for handwritten content covering text areas, it is necessary to restore any missing information based on the text context; for handwritten content covering graphic areas, it is necessary to restore any lines that may be affected based on the graphic structure.

[0136] In one implementation, a context-based prediction-based reconstruction method can be used to reconstruct the target region. Specifically, the reconstruction structure can receive: image features of the unprocessed region; the location of the target region; page layout information; content category information, etc. Then, based on the above information, the content that the target region should have is predicted. For example, for a text region covered by handwriting, the reconstruction structure can predict missing characters based on the relationship between adjacent characters; for a covered background region, a continuous background can be generated based on the surrounding texture; for a covered graphic region, missing lines can be recovered based on the graphic structure.

[0137] In one implementation, a generative image restoration model can be used to achieve the above restoration process. The generative image restoration model may include: an encoding structure for extracting features from the input image; a contextual association structure for analyzing the relationship between the target region and its surrounding regions; and a decoding structure for generating the restoration result. The above model structure is merely one implementation method, and this application does not limit the specific model used.

[0138] In other implementations, rule-based image inpainting algorithms, texture matching-based restoration methods, diffusion model-based generative restoration methods, and visual language model-based content restoration methods can also be employed. As long as the restored content can be determined based on non-target information regions while maintaining the original information expression capability, it can be used to achieve the information content reconstruction processing of this application.

[0139] In one implementation, for the multiple images to be processed determined in step S2, the above-described information content reconstruction process can be performed separately to obtain multiple reconstructed images. Using multiple images to be processed separately for reconstruction has the following advantages: although different images have all undergone defect screening, their local information may still differ. For example: the text on the first image is clear, but some illustration areas are occluded; the illustrations in the second image are complete, but some text areas are slightly blurred; the overall quality of the third image is good, but the user's handwriting covers a large area. By performing information content reconstruction separately on the above different images, multiple different reconstruction results can be obtained. Subsequently, through an information retention evaluation process, the multiple reconstruction results can be compared to select the most suitable data as the standard image. Therefore, this application does not require determining a single best image at the input stage, but allows the retention of multiple candidate results with high processing value, and further screening can be achieved through subsequent evaluation processes.

[0140] In one implementation, the information content reconstruction model can be obtained through sample training. Specifically, a training sample set can be constructed, which may include: the original information carrier image; the image to be processed containing non-target information; annotations of non-target information regions; and the reconstructed target image. During training, the image to be processed can be input into the information content reconstruction model, the model outputs the reconstruction result, and the model parameters are adjusted based on the difference between the output result and the target image. For example, a region segmentation loss constraint model can be used to accurately identify non-target information regions; a reconstruction loss constraint can be used to restore the consistency between the result and the target content; and a structure preservation loss can be used to constrain the text, formulas, and graphic structures to prevent abnormal changes.

[0141] It should be noted that the above-described model structure, training data, and optimization method are only used to illustrate one implementation method of this application. The focus of this application is not on the specific network structure or training method used, but on determining the target processing region based on the content type in the image to be processed, and performing information content reconstruction processing that can maintain the original information expression ability.

[0142] Through the above methods, this application can remove user-added or non-target information while preserving the original content of the information carrier, and generate a reconstructed image that better meets the requirements of standard resources through content restoration. Compared with traditional methods that directly delete abnormal areas, this application can make a comprehensive judgment by combining text structure, formula structure, and graphic structure information, reducing the risk of erroneous deletion of effective information and improving the reliability of the reconstruction results.

[0143] Meanwhile, by further evaluating the reconstructed image execution information in subsequent steps, the problem of judging the reconstruction quality solely based on visual effects can be avoided, ensuring that the final generated standard image not only has a more standardized appearance but also retains the ability to express the original information.

[0144] S4. Determine the information preservation evaluation result based on the information preservation relationship between the reconstructed image and the image to be processed.

[0145] In one implementation, after obtaining the reconstructed image corresponding to the image to be processed, the information preservation evaluation result can be further determined based on the information preservation relationship between the reconstructed image and the image to be processed.

[0146] It should be noted that the information content reconstruction process in step S3 mainly focuses on how to remove or restore non-target information in the image to be processed, while step S4 mainly focuses on whether the effective information in the original information carrier is preserved after reconstruction processing. In other words, step S3 solves the problem of "how to process image content", while step S4 solves the problem of "whether the processed image still retains the original information expression ability".

[0147] In practical applications, judging the quality of a processing result solely based on the visual effect of the reconstructed image has significant limitations. For example, an image after removing handwritten content may appear cleaner, but some printed text may have been incorrectly deleted; another image may have an overall visual effect close to the original, but certain formula symbols, illustration lines, or structural relationships may have changed. Therefore, it is necessary to establish an evaluation mechanism for information preservation capabilities, rather than judging solely based on the degree of pixel change or human visual effects.

[0148] For example, in textbook page processing, if a student has annotated the text area of ​​the question, the information content reconstruction model might correctly remove some of the annotations, but mistakenly delete characters in the question stem. In this case, the reconstructed image may have fewer black pixels than the original image, but its actual information value decreases. If only pixel differences or image similarity are used for evaluation, this problem might not be detected. Therefore, in one implementation, an information preservation evaluation can be performed on the reconstructed image based on the changes in information before and after reconstruction. The information preservation evaluation result is used to represent the degree to which the reconstructed image retains the original effective information relative to the original image.

[0149] The information preservation relationships can include: information expression area preservation relationships; information structure preservation relationships; information association relationship preservation relationships; and spatial location preservation relationships. By comprehensively analyzing the above preservation relationships, it is possible to more accurately determine whether the information content reconstruction process has resulted in the loss of effective information.

[0150] In one implementation, information representation region analysis can be performed on both the image to be processed and the reconstructed image to identify regions in the image that have information representation value. Specifically, not all regions of an information carrier image typically have the same information value. For example, on a textbook page, titles, title text, formulas, illustrations, and tables usually carry the main information, while blank areas at the page edges, decorative areas, or background areas have lower information value. If a uniform method is used to evaluate all regions, the evaluation results may not accurately reflect the actual information retention. For instance, although a certain region may have significant pixel changes, if that region is a blank background, its impact on information representation is minimal; another region may have only minor character changes, but because it is located in the title text area, the entire title information may be incomprehensible. Therefore, in one implementation, information region recognition processing can be performed on the image to be processed, dividing the image into multiple different types of information representation regions.

[0151] The information expression area may include: (1) a text expression area; (2) a formula expression area; (3) a graphic expression area; (4) a table expression area; and (5) a background area. The text expression area represents the area that expresses information through character arrangement, such as title text, explanatory text, and label text; the formula expression area represents the area that expresses information through combinations of mathematical symbols; the graphic expression area represents the area that expresses information through spatial structure, such as diagrams, flowcharts, and structural diagrams; the table expression area represents the area that expresses information through row and column relationships; and the background area represents the area that does not directly carry semantic information.

[0152] In one implementation, an image understanding model can be used to perform region classification processing on the image to be processed. The image understanding model may include: an image feature extraction structure; a region classification structure; and a semantic association analysis structure. Specifically, the image feature extraction structure is used to extract visual features from the image to be processed; the region classification structure is used to determine different region categories based on visual features; and the semantic association analysis structure is used to determine the information type corresponding to a region by combining contextual relationships.

[0153] For example, if a region is detected to contain regularly arranged characters and form a continuous text structure with adjacent characters, it can be determined that the region belongs to the text expression region; if the region is detected to contain fractional structures, subscript-superscript relationships, or special mathematical symbols, it can be determined that the region belongs to the formula expression region; if the region is detected to contain continuous edges, node connections, or closed structures, it can be determined that the region belongs to the graphic expression region.

[0154] It should be noted that the above region classification process is only used to illustrate one implementation method, and this application does not limit the specific model or algorithm used. In other embodiments, traditional image analysis methods, OCR analysis methods, layout analysis methods, or multimodal understanding models can also be used to achieve information expression region recognition.

[0155] In one implementation, after identifying different information expression regions, the relationship between the information structure preservation between the reconstructed image and the image to be processed can be further analyzed. Specifically, different types of information have different expression structures, therefore a unified evaluation index cannot be used for judgment.

[0156] For example, the core of textual information lies in the content and arrangement of characters; for formulaic information, the core lies in the combination of symbols and mathematical structure; and for graphical information, the core lies in spatial connections and overall topological structure. Therefore, it is necessary to establish corresponding evaluation methods based on different information types.

[0157] In one implementation, for text representation regions, the degree of text information preservation can be determined by comparing the changes in text structure before and after reconstruction. Specifically, text feature extraction can be performed on the text regions in the image to be processed and the reconstructed image respectively to obtain corresponding text feature information. The text feature information may include: character features; row and column arrangement features; text region layout features; and relationships between characters. Subsequently, the feature differences of corresponding text regions in the two images can be compared.

[0158] For example, if the text region in the reconstructed image maintains a consistent number of characters, consistent character arrangement, basically consistent line spacing, and no significant changes in character structure, then the region can be considered to have a high degree of text preservation. Conversely, if there are missing characters, character replacements, line structure disruptions, or the disappearance of local text regions, then the region is considered to have lost text information.

[0159] In one implementation, the evaluation method for formula regions can differ from that for ordinary text regions. This is because formula information depends not only on individual characters but also on the positional relationships between symbols. Examples include the positional relationship between numerators and denominators, superscripts and subscripts, the coverage of square roots, and operator connections. Therefore, even if the number of characters in the formula region is not significantly reduced, changes in the spatial structure can alter the meaning of the formula. For instance, if an exponent symbol in the original formula is mistakenly deleted, the remaining characters may still exist, but the formula's semantics have changed.

[0160] Therefore, in one implementation, structural features corresponding to the formula region can be extracted, and the structural consistency before and after reconstruction can be compared. These structural features may include: symbol category, symbol position, symbol connection relationship, and hierarchical structure relationship. Through this method, it can be determined whether the reconstruction of information content affects the formula expression.

[0161] In one implementation, the degree of preservation for a graphical representation area can be determined based on changes in the graphical structure. Unlike text, graphical information typically expresses meaning through spatial relationships. For example, a line connects two nodes; an arrow indicates the direction of a flow; a closed region represents a specific structure. Therefore, graphical evaluation cannot simply compare pixel differences but must consider structural relationships. For instance, in a flowchart, if a connecting line is removed, although the overall image change area is small, it may lead to an incomprehensible flow relationship; in geometry, if a boundary line is missing, it may cause a change in the properties of the graphic.

[0162] Therefore, in one implementation, structural features corresponding to the graphic region can be extracted, including edge information, line segment connectivity, node relationships, and region topological relationships. Subsequently, by comparing the structural differences between the reconstructed image and the image to be processed, the graphic preservation evaluation result is determined.

[0163] In one implementation, in addition to evaluating whether different information expression regions themselves remain intact, it is also possible to further evaluate whether the information relationships between different information expression regions are maintained. This is because, in actual information carriers, information content does not exist completely independently; there are usually semantic or structural relationships between different regions. For example, there is a content correspondence between the title text and the corresponding illustration on a textbook page, a structural relationship between the symbols and superscripts / subscripts in a formula, and a logical relationship between the nodes and connecting lines in a flowchart. If the information content reconstruction process only maintains the content of local regions while destroying the relationships between regions, the final image may still fail to correctly express the original information.

[0164] For example, on a math problem page, the problem statement, formulas, and geometric figures all describe the same problem. If the problem statement is retained during the reconstruction process, but key boundary lines in the corresponding figures are deleted, the overall information expression will change, even though the text area remains intact. As another example, in technical documentation, there is a relationship between the component names in a structural diagram and the positions of their corresponding figures. If the text labels are retained during the reconstruction process, but the positional relationship between the labels and the figures is changed, it may lead to errors in subsequent understanding of the model.

[0165] Therefore, in one implementation, the information association preservation result can be determined based on the regional relationships in the image to be processed and the reconstructed image. Specifically, the relationships between different information expression regions can be determined first. These relationships can include: spatial proximity; semantic correspondence; structural connection; and content referencing. Spatial proximity represents the positional relationship between different information regions, such as the positional relationship between the title text and the corresponding option region; semantic correspondence represents the association of content expressed between different regions, such as the relationship between text descriptions and corresponding illustrations; structural connection represents the connection between graphic elements, such as the relationship between process nodes and arrows; and content referencing represents the information referencing relationship between different regions, such as the relationship between formula numbers and formula bodies.

[0166] In one implementation, information regions in the image to be processed can be constructed as an information association graph. For example, each information representation region is treated as a node; the relationships between different regions are treated as edges; and the weights of the edges represent the strength of the relationships. Subsequently, the information association graphs corresponding to the image to be processed and the reconstructed image can be constructed separately, and the changes between the two association graphs can be compared.

[0167] For example, if text nodes and graphic nodes still exist in the reconstructed image, and the edges connecting them haven't changed significantly, it indicates that the information associations are well maintained. Conversely, if a text region disappears, a graphic node is missing, or the original connections are broken, it means the reconstruction process affected the expression of information associations. This evaluation method avoids the problem of traditional image evaluation methods focusing only on local content, making the evaluation results more aligned with the actual expressive needs of the information carrier.

[0168] In a preferred embodiment, in order to further improve the accuracy of information retention evaluation, the defect state characterization information obtained in step S1 can be used to impose spatial prior constraints on the information retention evaluation process.

[0169] Specifically, in step S1, the defect status information corresponding to the image to be processed has been obtained, including the defect probability matrix corresponding to different defect types. This information reflects the reliability of information in each spatial region of the image before it enters the information content reconstruction process. However, ignoring this prior information during the information preservation evaluation process may lead to incorrect judgments. For example, if a certain region in the image to be processed is already invisible due to occlusion defects, a difference will arise between this region and the original image after information content reconstruction. If the evaluation model is unaware that this region was originally occluded, it may assume that the difference is caused by the information content reconstruction process, thereby lowering the evaluation of the reconstruction result. In reality, this difference originates from the defects in the input image itself, rather than from information loss during the reconstruction process.

[0170] Therefore, in one implementation, defect state characterization information can be incorporated into the information retention evaluation process, enabling the evaluation model to distinguish whether the information change is caused by defects in the original data acquisition or by information content reconstruction. This approach can improve the accuracy of the evaluation results.

[0171] In a preferred embodiment, a spatial prior gating mechanism can be used to achieve information-preserving evaluation adjustment based on defect state. Specifically, the spatial prior gating mechanism is used to generate spatial gating weights based on the defect probability matrix, and to adjust the influence of features corresponding to different spatial locations in the evaluation process using the spatial gating weights.

[0172] Traditional image comparison methods typically assume that all regions in an image have the same level of reliability. For example, when comparing the differences between the image to be processed and the reconstructed image, the model assumes that changes in each region have the same evaluative significance. However, in practical applications, the reliability of information in different regions may vary significantly. For instance, the original image in region A may be clear, without occlusion or blurring. The original image in region B may have significant occlusion, with some content not visible. If the reconstructed image and the image to be processed differ in both regions, the difference in region A is more likely caused by the information content reconstruction process, while the difference in region B may originate from defects in the original acquisition. Therefore, it is necessary to enable the evaluation model to have spatial awareness.

[0173] In one implementation, spatial gating weights can be generated based on the defect probability matrix.

[0174] For example, for the occlusion probability matrix This can generate corresponding spatial reliability weights:

[0175] in: This represents the occlusion probability corresponding to the pixel spatial location (x, y). This indicates the reliability of information corresponding to the pixel spatial location (x, y). When the probability of occlusion in a certain region is high, the corresponding weight is low; when a certain region is not occluded, the corresponding weight is high.

[0176] Similarly, for cases involving multiple defect types, the comprehensive spatial reliability weight can be obtained by combining multiple defect probability matrices. The calculation method is similar to the formulas described above, and will not be elaborated upon here.

[0177] In one implementation, spatial gating weights can also be applied to the feature fusion process.

[0178] For example, an information preservation evaluation model can simultaneously receive: features of the image to be processed; features of the reconstructed image; and features of the defect state. The features of the image to be processed provide a reference for information before reconstruction; the features of the reconstructed image provide information about the results after processing; and the features of the defect state provide constraints on spatial reliability.

[0179] During feature fusion, spatial gating weights can be used to adjust the contribution levels of features from different regions. For high-reliability regions, the feature contribution of the corresponding region can be increased, enhancing change detection capabilities. For low-reliability regions, the feature contribution of the corresponding region can be decreased, reducing the risk of misjudgment. Through these methods, the evaluation model can focus more on the information changes truly caused by the reconstruction process.

[0180] By introducing a spatial prior gating mechanism, this application can use the defect state information obtained in the previous stage to guide the evaluation process in the subsequent stage, so that the evaluation model can not only determine "whether the image has changed", but also "whether the change has evaluation significance".

[0181] Compared with traditional image comparison methods that do not consider input reliability, this application has the following advantages: First, it can reduce the interference of defects in the original data acquisition on the evaluation results. For example, information differences caused by defects such as occlusion and blurring will not be simply considered as reconstruction failure.

[0182] Second, it can improve the ability to detect changes in key areas. For example, greater attention can be paid to information loss in clear text areas or complete graphic areas.

[0183] Third, it can form an information loop between the preceding and following stages. The results of the defect analysis in the preceding stage are no longer just used for input screening, but are further involved in the subsequent evaluation process, improving the overall reliability of the processing.

[0184] In one implementation, in order to more accurately evaluate the information retention before and after information content reconstruction processing, a print and illustration integrity comparison model can be constructed, and the print and illustration integrity comparison model can be used to analyze the image to be processed and the reconstructed image.

[0185] It should be noted that in information carrier images, printed text and illustrations typically serve different informational functions, and they differ significantly in structural characteristics, variations, and evaluation methods. Therefore, using a uniform evaluation method to analyze all areas may not accurately reflect the degree of preservation of different types of information.

[0186] For example, for printed text areas, the main focus is on whether characters are missing, whether the text arrangement has changed, and whether the text semantics are preserved; while for illustration areas, the main focus is on whether the line structure, spatial relationships, and graphic topological relationships are preserved. If only pixel-level differences are used for judgment, it may be impossible to distinguish between slight visual changes and actual information loss.

[0187] Therefore, in one implementation, a print body and illustration integrity comparison model can be used to evaluate print body information and illustration information separately. The print body and illustration integrity comparison model may include: a first input branch; a second input branch; a feature fusion structure; a text integrity evaluation branch; and an illustration integrity evaluation branch. The first input branch is used to input the image to be processed, and the second input branch is used to input the reconstructed image. By simultaneously inputting the images before and after reconstruction, the model can obtain a reference for information changes, rather than judging content integrity based solely on a single image.

[0188] For example, if only the reconstructed image is input, the model cannot determine whether certain areas on the current page originally contained content, nor can it determine whether the missing content in a certain area was caused by the reconstruction process; however, by simultaneously inputting the image to be processed, the changes in the corresponding areas before and after processing can be clearly compared, thereby improving the accuracy of the evaluation.

[0189] In one implementation, the print and illustration integrity comparison model may include a text integrity evaluation branch. This branch evaluates whether the printed text in the reconstructed image remains intact. Specifically, the text integrity evaluation branch can determine text region changes based on fusion features and output a result indicating text information preservation.

[0190] In one implementation, text regions in the image to be processed and the reconstructed image can be located first. Subsequently, feature comparisons are performed on the corresponding text regions. The evaluation may include: (1) Character existence evaluation, used to determine whether the original characters have disappeared during the reconstruction process. For example, the original image contains: "x²+y²=z²". If the superscript "²" is missing in the reconstructed image, it indicates that text information loss has occurred.

[0191] (2) Character structure evaluation, used to determine whether the character form has undergone abnormal changes. For example, the number "8" is restored to a structure similar to "3", and the letter "m" is restored to "n". Although the character area still exists, the semantics may have changed.

[0192] (3) Evaluation of text arrangement relationship, used to determine whether line spacing, character spacing, text direction and area layout are maintained intact. For example, during the page reconstruction process, if a piece of text is moved incorrectly, even if the characters are intact, it may affect the understanding of the page.

[0193] In one implementation, a text loss region can be generated based on the text region prediction results. When a region is determined to be a text loss region, it can be considered that the information content reconstruction process has affected the integrity of the printed text.

[0194] Specifically, the text loss region is used to represent the areas in the image to be processed where the original text content was not fully preserved after information content reconstruction. This application can determine whether the character content, character structure, and text arrangement have changed by comparing the text region features in the image to be processed with the corresponding text region features in the reconstructed image.

[0195] For example, if a text region in the image to be processed contains complete characters, but the corresponding region in the reconstructed image shows missing characters, abnormal character shapes, changes in character connections, or disruption of text arrangement, then the region can be considered to have suffered text information loss. Conversely, if the corresponding text region in the reconstructed image retains its original character structure and arrangement, then the region can be considered to have not suffered significant text loss.

[0196] Furthermore, the text loss can be comprehensively assessed based on the importance of information corresponding to different text regions. For example, for information regions carrying the main semantic meaning, such as the title text and key explanatory text, even a small number of character changes may affect the overall information expression, thus increasing the evaluation attention given to these regions; while for page edges, page numbers, or decorative text regions, even if there are some changes, the impact on the overall information expression is small, thus reducing the evaluation impact of these regions.

[0197] By using the above methods, we can determine whether the information content reconstruction process has caused a loss of printed information from the perspective of text content preservation, rather than relying solely on the degree of change in image pixels, thereby improving the accuracy of the information preservation evaluation results.

[0198] In one implementation, the print and illustration integrity comparison model may further include an illustration integrity evaluation branch. This branch evaluates whether the illustration structure in the reconstructed image remains intact. Unlike textual information, illustration information typically relies on spatial structural representations. For example, geometry relies on boundaries and connections, flowcharts on nodes and lines, and technical structure diagrams on component layouts. Therefore, illustration integrity evaluation cannot rely solely on pixel differences.

[0199] In one implementation, structural features corresponding to the inset region can be extracted. These structural features may include edge features, line segment directions, node positions, region connectivity, and topological structure. Subsequently, changes in structural features between the original image and the reconstructed image can be compared. For example, if the original image contains node A—node B—node C, but the connection between node A and node B disappears in the reconstructed image, it indicates a loss of graphic structure.

[0200] In one implementation, an inset loss region can be generated. Specifically, the inset loss region represents areas in the image to be processed where the original inset information was not fully preserved after information content reconstruction. Since inset information typically relies on spatial structure, edge relationships, and connectivity between elements, the evaluation of the inset region cannot be based solely on the degree of pixel change; further analysis is needed to determine whether the graphic structure has changed.

[0201] In one implementation, it can be determined whether the illustration content has been lost by comparing the features of the illustration region in the image to be processed with the features of the corresponding illustration region in the reconstructed image. For example, when the illustration region in the reconstructed image shows missing key boundary lines, changes in graphic contours, broken connections, missing structural nodes, or abnormal changes in local areas, it can be considered that the corresponding region has lost illustration information.

[0202] For example, in a diagram containing flow relationships, if connecting lines indicating flow directions are deleted during reconstruction, although the overall changed area of ​​the image is small, the meaning of the illustration changes because the relationship between the graphics is disrupted. Therefore, this area can be identified as an illustration loss area. Similarly, for geometric figures, if reconstruction causes a boundary line to disappear or the closure relationship of the figure to be broken, it may affect the subsequent understanding of the figure's properties; this area can also be considered to have illustration information loss.

[0203] Furthermore, the integrity evaluation result of the illustration can be determined based on the location, extent, and information importance of the lost illustration area. For illustration areas that carry major structural information, such as key component areas in technical structure diagrams, key connection areas in flowcharts, or key boundary areas in mathematical graphics, the influence of the corresponding areas in the integrity evaluation can be increased; for decorative patterns or auxiliary areas that do not affect the expression of information, even if there are some changes, their impact on the evaluation result can be reduced.

[0204] By using the above methods, we can determine whether the information content reconstruction process affects the original graphic's expressive ability from the perspective of illustration structure preservation, avoiding the evaluation of illustration quality solely based on visual similarity or pixel differences, thereby improving the accuracy of information preservation evaluation results.

[0205] In one implementation, the overall information preservation evaluation result can be determined based on the text preservation evaluation value and the illustration preservation evaluation value. Specifically, text information and illustration information together constitute the main effective information content in the information carrier image. The text preservation evaluation value reflects the information changes in the text region before and after the information content reconstruction process, mainly focusing on whether character integrity, text arrangement relationships, and semantic expression are preserved. The illustration preservation evaluation value reflects the information changes in the illustration region before and after the reconstruction process, mainly focusing on whether graphic structure, edge relationships, and regional association relationships are preserved. When determining the overall information preservation evaluation result, the information preservation status of the text region and the illustration region can be comprehensively considered to obtain an evaluation result that reflects the degree of impact of the entire reconstruction process.

[0206] In one implementation, the text retention evaluation results and illustration retention evaluation results can be comprehensively analyzed based on the importance of different information types in the current application scenario. For example, in application scenarios such as textbook pages and test paper pages, text content usually carries the main semantic information, thus increasing the impact of text retention evaluation results on the overall information retention evaluation; in information carriers containing a large number of structural diagrams, flowcharts, or schematic diagrams, illustration structures may play a major expressive role, thus increasing the impact of illustration retention evaluation results.

[0207] Furthermore, the degree of influence of the evaluation results can be determined based on the information value of different regions. For example, for regions containing key formulas, core textual descriptions, or key structural relationships, even small changes in local areas may lead to changes in the overall information expression; therefore, the preservation of information in these regions should be a key focus. For auxiliary explanatory areas or non-critical decorative areas, their impact on the overall evaluation results can be reduced.

[0208] In another implementation, the overall information preservation evaluation result can be determined based on the correlation between the text preservation evaluation value and the illustration preservation evaluation value. For example, when both the text region and the illustration region are well preserved, the reconstructed image can be considered to have maintained a high degree of information integrity; when any important information region suffers significant loss, even if other regions are well preserved, the overall information preservation evaluation result can be reduced. Through this method, the impact of information content reconstruction processing on the original information can be comprehensively evaluated from multiple information dimensions, avoiding judgment of reconstruction effectiveness based solely on a single type of information, and improving the accuracy and reliability of the overall information preservation evaluation result.

[0209] It should be noted that this application does not limit the specific method of integrating the text retention evaluation value and the illustration retention evaluation value. In different implementations, weighted analysis, rule judgment, model evaluation, or other data fusion methods can be used to determine the overall information retention evaluation result according to actual application needs. As long as the degree of information retention can be evaluated based on the changes in text information and illustration information before and after reconstruction, it falls within the scope of this application.

[0210] In one implementation, to enable the printout and illustration integrity comparison model to accurately learn the loss of different types of information, a joint loss function can be designed during model training. Since the model in this application needs to determine not only whether the image has changed, but also whether different types of information are preserved, using a single classification loss or a single pixel loss may not meet the training requirements. Therefore, in one implementation, a joint loss function can be used:

[0211] Where L represents the total training loss of the model, Indicates loss of text integrity. Indicates loss of illustration integrity; Indicates structural retention loss. , , These represent the corresponding loss weights.

[0212] Specifically: text integrity loss is used to constrain the model to accurately identify changes in printed text; illustration integrity loss is used to constrain the model to accurately identify changes in graphic structure; and structure preservation loss is used to constrain the overall layout of the page and the preservation of information relationships.

[0213] In one implementation, the aforementioned joint loss function can be used to train the printout and illustration integrity comparison model, enabling the model to learn the information preservation relationship between the image to be processed and the reconstructed image. Specifically, during model training, a training sample set can be constructed. The training sample set may include: the original information carrier image; the image to be processed containing non-target information; the corresponding reconstructed information content image; text information preservation annotations; illustration information preservation annotations; and structural relationship preservation annotations. The original information carrier image can represent the ideal state of the information carrier; the image to be processed containing non-target information can represent the image uploaded by the user during actual acquisition; and the reconstructed information content image can represent the result after reconstruction processing.

[0214] During training, the image to be processed and its corresponding reconstructed image can be input into the text and illustration integrity comparison model. This allows the model to extract the difference features between the two and calculate the difference between the predicted and labeled results based on the joint loss function. For example, for the text integrity branch, the corresponding training optimization direction can be determined based on the difference between the model's output text preservation prediction and the actual text preservation label.

[0215] Specifically, during model training, text regions in the training samples can be pre-labeled to determine the text content state that should be preserved after the original image undergoes information content reconstruction. After receiving the image to be processed and the corresponding reconstructed image, the model can analyze the changes in text regions between the two and output the text preservation prediction result.

[0216] Subsequently, the text preservation prediction results can be compared with the pre-set text preservation annotation results. When there is a large difference between the model prediction results and the actual annotation results, it indicates that the model's judgment on the preservation of text regions is biased. At this time, the model parameters can be adjusted through the training process to allow the model to gradually learn the correspondence between changes in text content and information loss. When the prediction results are close to the actual annotation results, it indicates that the model can accurately judge whether information loss has occurred in the text region.

[0217] By employing the above methods, the text integrity branch can focus on the changes in text content before and after reconstruction, rather than simply judging the degree of image change based on pixel differences. For example, for minor changes in font color or background texture that do not affect text expression, the model can reduce their impact on the evaluation results; while for situations that affect semantic expression, such as missing characters, changes in character structure, or disruption of text arrangement, the model can increase its attention to the corresponding regions.

[0218] It should be noted that the above training process is only used to enable the text integrity branch to have text preservation evaluation capabilities. The specific training methods, optimization strategies, and model parameter adjustment methods used are not the focus of this application. This application focuses on determining the text preservation evaluation result by analyzing the information preservation of text regions after the information content reconstruction is completed, and further combining the illustration preservation evaluation result to determine the overall information preservation evaluation result.

[0219] In other implementations, different model training methods or rule-based judgment methods that do not require training can also be used to evaluate text integrity. For example, the preservation of text before and after reconstruction can be determined by comparing text recognition results, analyzing character structure, or matching text region features. As long as it is possible to determine whether text information is preserved based on changes in text regions, it falls within the scope of the technical solution of this application.

[0220] In one implementation, the use of the above-mentioned joint loss function has the following technical effects.

[0221] First, it avoids evaluation bias caused by a single visual loss. For example, when using only pixel difference loss, the model might consider significant changes in the background area or obvious changes in page color as serious information loss. However, in reality, changes in background color may not affect information expression, while the absence of a character or the breakage of a key connecting line could lead to serious information loss. By introducing text loss and illustration loss, the model can focus more on areas that truly have informational value.

[0222] Secondly, it can improve the targeted evaluation ability of different types of information. Since text and illustrations have different modes of expression, it is difficult to accurately describe their variations using a uniform evaluation index. By setting separate text integrity losses and illustration integrity losses, the model can learn corresponding evaluation features for different information types.

[0223] Secondly, it can improve the evaluation stability in complex page scenarios. In practical applications, the same page may simultaneously contain titles, body text, formulas, illustrations, and tables. Different areas are affected by reconstruction processing to varying degrees. The joint loss function can comprehensively consider the changes in different areas, making the final evaluation results more in line with the needs of practical applications.

[0224] It should be noted that the information preservation evaluation method based on deep learning models and joint loss functions described above is only one implementation method. In other implementations, non-deep learning methods can also be used to achieve information preservation evaluation. For example, rule-based evaluation methods can be used. Specifically, the degree of information preservation of the reconstructed image can be determined by comparing OCR recognition results, counting character counts, comparing page layouts, and analyzing graphic structures. Therefore, this application does not limit the specific evaluation algorithm used, but emphasizes using the information correlation between the images before and after reconstruction to determine the degree of information preservation.

[0225] In one implementation, after obtaining the text retention evaluation results, illustration retention evaluation results, and information association retention evaluation results, a comprehensive information retention evaluation result can be further determined.

[0226] For example:

[0227] in, This indicates that the comprehensive information maintains the evaluation results. This indicates that the text retains the evaluation results. The illustrations indicate that the evaluation results are preserved. This indicates that the information is related to maintain the evaluation results. , , This indicates the corresponding weight.

[0228] The weighting parameters can be adjusted according to the application scenario. For example, in the scenario of constructing an educational question bank, since the question stem and formulas usually determine the semantics of the question, the evaluation weight of text and formula relevance can be increased; in the scenario of digitizing engineering data, since structural diagrams may carry key technical information, the evaluation weight of illustration structure can be increased. Furthermore, the comprehensive information retention evaluation result can be combined with the information quality evaluation result in step S2 for the final standard image selection. Through the above steps, this application can perform a more accurate information retention evaluation on the reconstruction result after the information content reconstruction is completed.

[0229] Compared with the prior art, this application has at least the following effects: First, this application does not simply compare pixel changes before and after reconstruction, but analyzes whether text, formulas, illustrations, and regional relationships are preserved from the perspective of information expression, thus enabling a more accurate evaluation of reconstruction quality. Second, this application utilizes defect state information obtained in the previous stage to establish spatial prior constraints, enabling the evaluation process to distinguish between original acquisition defects and information loss caused by content reconstruction, thereby reducing the probability of mis-evaluation. Third, this application processes different types of information through text preservation evaluation and illustration preservation evaluation respectively, making the evaluation results more consistent with complex information carrier scenarios. Fourth, this application, through a joint evaluation mechanism, ensures that the final evaluation result simultaneously considers local content preservation and overall structural preservation, improving the reliability of standard image generation.

[0230] S5. Based on the information, determine the target standard image according to the evaluation results, and store the target standard image in the database.

[0231] In one implementation, after obtaining information preservation evaluation results corresponding to multiple reconstructed images, a target standard image can be further determined based on the information preservation evaluation results, and the target standard image can be stored in a database.

[0232] It should be noted that steps S1 to S4 respectively complete: defect state analysis of candidate information carrier images; image selection; information content reconstruction; and information preservation evaluation of the reconstruction results. However, selecting the final image solely based on the information preservation evaluation results may still have certain limitations. For example, although a certain reconstructed image retains a high proportion of the original information, its overall visual quality may not meet the requirements for database use due to slight blurring in the original input image; another reconstructed image, although visually better, suffers from loss of some text regions.

[0233] Therefore, in one implementation, the information preservation evaluation results and image quality evaluation results can be comprehensively considered to further evaluate multiple reconstructed images, thereby determining the final target standard image. The information preservation evaluation results reflect whether the original information is preserved, whether the text content is complete, whether the illustration structure is complete, and whether the information relationships are maintained. The image quality evaluation results reflect the image clarity, page integrity, visual usability, and image standardization. By considering these factors simultaneously, selection bias caused by a single evaluation indicator can be avoided.

[0234] In one implementation, the comprehensive evaluation result can be determined based on multiple evaluation indicators.

[0235] For example, the overall score can be determined as follows:

[0236] in: This represents the comprehensive evaluation result corresponding to the reconstructed image; This indicates that the information retains the evaluation results; This indicates the image quality evaluation result; This indicates the text recognition evaluation result; This indicates the corresponding weight.

[0237] Among these evaluation methods, the information preservation evaluation result measures whether the reconstruction process causes a loss of original valid information; the image quality evaluation result measures whether the image meets the standard resource storage requirements; and the text recognition evaluation result is used to further determine whether the text content in the image is suitable for subsequent intelligent processing. Through this comprehensive evaluation, a judgment can be made on candidate reconstruction results from two dimensions: "content correctness" and "visual usability."

[0238] In one implementation, the weights of different evaluation indicators can be adjusted according to the application scenario. For example, in the scenario of building an educational resource database, since the text of the questions, formulas, and illustrations are the core content, the weight of the information retention evaluation result can be increased. In the scenario of displaying image resources, since the user's visual experience is more important, the weight of the image quality evaluation can be appropriately increased. In the scenario of digitizing technical documents, since structural diagrams and parameter information are important, the proportion of graphic structure evaluation in the information retention evaluation can be increased. Therefore, this application does not limit the fixed weights of each evaluation indicator, but can dynamically adjust them according to actual application needs.

[0239] In one implementation, the reconstructed image can be further evaluated using text recognition to determine whether it meets the requirements for subsequent text understanding. Specifically, a text recognition model can be used to recognize the reconstructed image, and the text quality evaluation result can be determined based on the recognition result. For example, the number of recognizable characters, the accuracy of character recognition, the completeness of text regions, and the regularity of text arrangement can be evaluated.

[0240] It's important to note that text recognition evaluation is not intended to replace information preservation evaluation, but rather serves as a supplementary evaluation metric. This is because text recognition models may not be able to accurately determine whether graphic structure is preserved; furthermore, while some visually high-quality images can be recognized, they may have lost some of their original structural information. Therefore, it's necessary to combine text recognition with information preservation evaluation for a comprehensive assessment.

[0241] In one implementation, after obtaining comprehensive evaluation results for multiple reconstructed images, a target standard image can be determined based on these results. Specifically, the multiple reconstructed images can be sorted according to the comprehensive evaluation results, and the reconstructed image with the highest evaluation result can be selected as the target standard image. For example, the first reconstructed image has an information retention evaluation value of 0.95, an image quality evaluation value of 0.80, and a text recognition evaluation value of 0.96. The second reconstructed image has an information retention evaluation value of 0.85, an image quality evaluation value of 0.98, and a text recognition evaluation value of 0.92. Based on the comprehensive scoring method, this application can select the first reconstructed image that is more suitable as a standard resource. This selection method avoids the problem of selecting an image solely based on visual effects or a single indicator.

[0242] In one implementation, after determining the target standard image, it can be stored in a database. Specifically, the database can simultaneously store the standard image file, image source information, information object identifier, quality evaluation results, information retention evaluation results, and processing parameters. By saving the above-mentioned associated information, subsequent image retrieval, content analysis, data tracking, and model optimization can be supported. For example, in an educational resource database, textbook chapter information, page numbers, image quality scores, and text recognition results can be stored simultaneously. In this way, when a user queries a certain knowledge point, the corresponding standard image resource can be quickly located.

[0243] In one implementation, to improve the quality of database resources, the stored standard images can be continuously updated. For example, when new candidate information carrier images are subsequently obtained, defect status analysis, information quality evaluation, content reconstruction, and information retention evaluation can be re-executed. If the comprehensive evaluation result corresponding to the new candidate image is higher than that of the existing standard image, the new standard image can replace the original image. Through the above methods, the standard image resources in the database can be continuously optimized as data accumulates. Through the above steps, this application can automatically determine the final standard image based on multiple evaluation dimensions.

[0244] Compared with the prior art, this application has the following advantages: First, instead of relying on a single image quality metric for image selection, we comprehensively consider information preservation capability and visual quality to improve the reliability of standard images.

[0245] Second, by using information preservation evaluation as the core evaluation factor, we can avoid the problem of sacrificing original information in order to improve visual effects.

[0246] Third, by combining text recognition evaluation, it can be ensured that the generated images are not only suitable for human viewing, but also for subsequent intelligent processing.

[0247] Fourth, through the database storage mechanism, an automated standard image resource construction process can be formed.

[0248] Thus, steps S1 to S5 in the method embodiment have all been described. The overall process is as follows: First, multiple candidate information carrier images are acquired, and the problems existing in different candidate images are determined through defect status analysis; second, suitable images to be processed are selected based on the defect status and information quality evaluation results; third, information content reconstruction processing is performed on the images to be processed to generate reconstructed images after removing non-target information; then, information preservation evaluation is used to determine whether the reconstruction process affects the original valid information; finally, the target standard image is determined based on the comprehensive evaluation results and stored in the database. Through the above closed-loop processing method, this application can realize the automatic generation of standardized digital resources from original user-uploaded images in complex non-standard image acquisition environments, improving the efficiency and quality stability of image resource construction.

[0249] In a specific application scenario, the method of this application can be applied to the construction of standard image resources for educational materials. This application is used to automatically generate standard image resources that can be used for question bank construction, intelligent retrieval, and knowledge analysis from a large number of non-standardized textbook, test paper, or exercise book page images uploaded by users.

[0250] For example, an educational platform needs to build a math question bank. Since the platform initially lacks a large number of high-definition standard textbook pages, it allows teachers, students, or operators to upload images of printed textbook pages or test paper pages via mobile devices to quickly build digital educational resources using crowdsourcing. During the actual data collection process, for the same math problem or the same textbook page, the platform may receive multiple candidate images uploaded by users. For example, for question 8 on page 35 of a math workbook, this application received the following multiple candidate images: First candidate image A: The user took a picture of the page directly with their mobile phone. The page is complete and the text is clear. However, the student had written calculations in the lower right corner of the page, and some of the handwritten content covered the blank area.

[0251] Second candidate image B: The user took the photo at an angle with their mobile phone. There is obvious perspective distortion on the left side of the page, but the text area of ​​the title is basically intact. At the same time, there is a large area of ​​reflection on the page.

[0252] Third candidate image C: A scanned image uploaded by the user. It has a high overall resolution, but due to insufficient scanning area, part of the bottom area of ​​the page is missing, which contains the answer options for the question.

[0253] Fourth candidate image D: The user took the picture from a good angle, and the page is complete, but during the shooting process, the user's finger covered part of the formula area in the question.

[0254] For the aforementioned candidate images, if the highest resolution image is selected directly using traditional methods, candidate image C might be chosen. However, candidate image C suffers from page truncation, resulting in the loss of key content, making it unsuitable as a standard title resource. The method in this application first acquires the aforementioned candidate information carrier images and then performs defect state analysis. Specifically, this application uses a defect detection model to analyze candidate images A, B, C, and D respectively.

[0255] For candidate image A, the following defects were detected: low probability of truncated page; low probability of occlusion; high probability of handwritten area; high page integrity. For candidate image B, the following defects were detected: high probability of text distortion; high probability of uneven lighting; decreased reliability of local text areas. For candidate image C, the following defects were detected: high probability of truncated page; missing information area at the bottom of the page; defect impact concentrated in the answer area. For candidate image D, the following defects were detected: high probability of occlusion; occlusion area located in the formula expression area; decreased reliability of local information.

[0256] Furthermore, this application generates corresponding defect state characterization information based on the aforementioned defect detection results. For example, for candidate image D, this application not only records "the existence of an occlusion defect," but also further determines that "the occlusion region is located in the formula region, and the information value of this region is high." Therefore, this occlusion defect corresponds to a high influence weight. Subsequently, this application determines the information quality evaluation result based on the defect state characterization information corresponding to different candidate images.

[0257] For example, the reliability of the page integrity, text area reliability, formula area reliability, illustration area reliability, and defect spatial distribution can be comprehensively considered. After evaluation, it was determined that: candidate image A had the highest information quality evaluation result; candidate image B had a lower evaluation result due to perspective and lighting issues; candidate image C had the lowest evaluation result due to missing key areas; and candidate image D had a lower evaluation result due to occlusion of formula areas. Therefore, candidate image A was selected as the image to be processed.

[0258] Furthermore, this application performs information content reconstruction processing on the image A to be processed. Since candidate image A contains the calculation process written by the student, this application first performs region analysis on the page content. Through content region recognition, the following are identified: the title text region; the mathematical formula region; the page illustration region; and the student's handwriting region. Among them, a corresponding handwriting mask is generated for the student's handwriting region. Subsequently, the student's handwriting content is removed according to the handwriting mask, and restoration processing is performed using the surrounding page background information.

[0259] For example, for calculations located in blank areas, the background is restored based on the surrounding paper texture; for handwritten content covering partially printed text, this application determines which printed characters need to be retained based on text structure relationships; for handwritten marks near formula areas, this application incorporates a formula structure protection mechanism to prevent accidental deletion of formula symbols. After processing, a corresponding reconstructed image is generated. However, this application does not directly consider the reconstructed image to be qualified, but further performs information preservation evaluation. Specifically, this application inputs the image to be processed A and the reconstructed image into the information preservation evaluation model. Among them, the text evaluation branch is used to determine: whether the question text is complete; whether mathematical symbols are preserved; and whether the character arrangement relationship has changed. The illustration evaluation branch is used to determine: whether the geometric boundaries are preserved; and whether the graphic connection relationship has changed. At the same time, this application uses the defect state information corresponding to candidate image A to generate spatial gating weights. Since candidate image A does not have obvious occlusion areas, the spatial weights corresponding to the text area and formula area are relatively high. If a formula symbol changes during the reconstruction process, the model will focus on this change and reduce the information preservation evaluation result. Finally, the following results are obtained: text preservation evaluation result; illustration preservation evaluation result; and information association preservation evaluation result. And a comprehensive information preservation evaluation value is determined based on the above results. Subsequently, by combining the information preservation evaluation results, image quality evaluation results, and text recognition evaluation results, a comprehensive evaluation result is calculated for the candidate reconstructed image. If the comprehensive evaluation result meets the preset criteria, the reconstructed image is determined as the target standard image and stored in the educational resource database.

[0260] The stored standard images can be further used for automatic question recognition, knowledge point association analysis, search retrieval, intelligent Q&A, and teaching resource recommendation. As can be seen from the above specific application scenarios, this application does not simply perform handwriting removal on user-uploaded images. Instead, it first analyzes the inherent defects of different candidate images and selects the more suitable input image for processing. Then, it performs information content reconstruction and uses information preservation evaluation to determine whether the reconstruction process affects the original valid content. Finally, it generates standard image resources suitable for database storage and intelligent applications. Compared to directly processing single user-uploaded images, this application can reduce the impact of acquisition defects and processing errors on the final resource quality and improve the reliability of the automated standard image generation process.

[0261] Please see Figure 3 This application provides a standard image generation apparatus based on information preservation evaluation, comprising: The defect status analysis module is used to acquire multiple candidate information carrier images and determine defect status characterization information based on the image defect status corresponding to the candidate information carrier images. The information quality evaluation module is used to determine the information quality evaluation result corresponding to each of the candidate information carrier images based on the defect state characterization information, and to determine the image to be processed from multiple candidate information carrier images based on the information quality evaluation result; The information content reconstruction module is used to perform information content reconstruction processing on the image to be processed to obtain a reconstructed image, and determine the information preservation evaluation result based on the information preservation relationship between the reconstructed image and the image to be processed. The information preservation evaluation module is used to determine the information quality fit degree corresponding to the reconstructed image based on the information preservation evaluation result and the image quality evaluation result corresponding to the reconstructed image, and to determine the target processing image based on the information quality fit degree. A standard image generation unit is used to generate a standard image based on the target processed image and store the standard image in an image database.

[0262] like Figure 4 As shown, an electronic device according to an embodiment of this application includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the standard image generation method based on information preservation evaluation as described in any one of this applications.

[0263] like Figure 5 As shown, this is an embodiment of a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the standard image generation method based on information preservation evaluation as described in any one of this applications.

[0264] The above Figures 3-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A target image screening method based on information representation capability evaluation, characterized in that, include: Multiple candidate information carrier images are acquired, and defect state characterization information is determined based on the image defect state corresponding to the candidate information carrier images. Based on the defect state characterization information, the information quality evaluation result corresponding to each candidate information carrier image is determined, and the image to be processed is determined from multiple candidate information carrier images based on the information quality evaluation result; Information content reconstruction processing is performed on the image to be processed to obtain a reconstructed image, and the information preservation evaluation result is determined based on the information preservation relationship between the reconstructed image and the image to be processed. Based on the information preservation evaluation results and the image quality evaluation results corresponding to the reconstructed image, the information quality adaptation degree corresponding to the reconstructed image is determined, and the target processing image is determined based on the information quality adaptation degree. A standard image is generated based on the target image, and the standard image is stored in an image database.

2. The method according to claim 1, characterized in that, The step of determining the defect state characterization information based on the image defect state corresponding to the candidate information carrier image includes: Defect detection is performed on the candidate information carrier image to determine the defect type corresponding to the candidate information carrier image; The defect types include at least one of truncated defects, occlusions, blurring, text distortion, angular rotation, and uneven lighting.

3. The method according to claim 2, characterized in that, The step of determining the defect state characterization information corresponding to the candidate information carrier image includes: For the defect characterization information corresponding to different defect types, determine the corresponding defect confidence information and defect occurrence probability information; Based on the defect confidence information and the defect occurrence probability information, defect state characterization information corresponding to the candidate information carrier image is generated.

4. The method according to claim 3, characterized in that, The step of generating spatially distributed defect state characterization information based on the defect state characterization information includes: Based on the defect confidence information and the defect occurrence probability information, the defect impact characterization results corresponding to different spatial locations in the candidate information carrier image are determined; Spatial distributed defect state characterization information is generated based on the defect impact characterization results.

5. The method according to claim 4, characterized in that, After generating spatially distributed defect state characterization information based on the defect impact characterization results, the method further includes: The reliability of information corresponding to different spatial regions is determined based on the spatially distributed defect state characterization information. Spatial weight information is generated based on the reliability of the information.

6. The method according to claim 5, characterized in that, The step of adjusting the contribution of image features corresponding to different spatial regions in subsequent processing based on the spatial weight information includes: Obtain the image features corresponding to the candidate information carrier image; Based on the spatial weight information, the image feature adjustment weights corresponding to different spatial locations are determined; Based on the image feature adjustment weights, the image features corresponding to different spatial locations are adjusted, and the adjusted image features are fused to generate fused image features.

7. The method according to claim 6, characterized in that, The step of adjusting the contribution level of image features corresponding to different spatial locations based on the spatial weight information includes: Based on the spatially distributed defect state characterization information corresponding to different spatial locations, determine the information expression capability of the corresponding spatial location; The spatial weight information is adjusted according to the information expression capability to obtain updated spatial weight information; The image features corresponding to different spatial locations are weighted according to the updated spatial weight information to generate fused image features.

8. A target image screening device based on information representation capability evaluation, characterized in that, include: The defect status analysis module is used to acquire multiple candidate information carrier images and determine defect status characterization information based on the image defect status corresponding to the candidate information carrier images. The information quality evaluation module is used to determine the information quality evaluation result corresponding to each of the candidate information carrier images based on the defect state characterization information, and to determine the image to be processed from multiple candidate information carrier images based on the information quality evaluation result; The information content reconstruction module is used to perform information content reconstruction processing on the image to be processed to obtain a reconstructed image, and to determine the information preservation evaluation result based on the information preservation relationship between the reconstructed image and the image to be processed. The information preservation evaluation module is used to determine the information quality fit degree corresponding to the reconstructed image based on the information preservation evaluation result and the image quality evaluation result corresponding to the reconstructed image, and to determine the target processing image based on the information quality fit degree. A standard image generation unit is used to generate a standard image based on the target processed image and store the standard image in an image database.

9. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the target image screening method based on information representation capability evaluation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the target image screening method based on information representation capability evaluation as described in any one of claims 1 to 7.

Images