An image privacy content recognition and protection system

Abstract

The application relates to the technical field of data processing, and relates to an image privacy content recognition and protection system, which comprises an image analysis module, a feature matching module, a mask generation module, a grade division module, a privacy processing module and an image fusion module, is used for analyzing a to-be-processed image to obtain a candidate privacy object, extracting features of the candidate privacy object to generate an object feature vector, matching the object feature vector with a preset privacy feature library to obtain a matching confidence, generating a privacy area mask according to the object feature vector, dividing the privacy area into different sensitive grades and matching corresponding privacy processing strategies, performing differential privacy processing on an image area covered by the privacy area mask according to the privacy processing strategies to generate an intermediate image, and performing image fusion on an unprocessed area in the intermediate image and a processed privacy area to generate an output image; and the application can improve the efficiency of image privacy content recognition and protection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an image privacy content recognition and protection system. Background Technology

[0002] Existing technologies cannot accurately analyze images by combining semantic segmentation and contour detection, struggle to extract local texture features and quantize them with privacy feature libraries, and cannot automatically generate accurate privacy region masks, resulting in insufficient accuracy and completeness in privacy content recognition.

[0003] Existing technologies have significant shortcomings in the privacy protection processing stage. They cannot classify sensitivity levels based on matching confidence and match differentiated processing strategies, cannot reconstruct and fill content based on neighborhood texture features, and do not perform smooth transition and fusion between the processing area and the original image. This results in poor privacy protection, low image visual integrity, and overall processing efficiency that is difficult to meet the requirements. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides an image privacy content recognition and protection system, characterized in that the system includes an image parsing module, a feature matching module, a mask generation module, a level classification module, a privacy processing module, and an image fusion module, wherein: The image parsing module is used to perform preliminary parsing on the image to be processed to obtain candidate privacy objects of the image to be processed; The feature matching module is used to extract features from the candidate privacy object, generate an object feature vector of the candidate privacy object, and match the object feature vector with a preset privacy feature library to obtain the matching confidence of the candidate privacy object. The mask generation module is used to generate a privacy region mask for the candidate privacy object based on the object feature vector when the privacy attribute of the candidate privacy object is determined to be private. The level division module is used to divide the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence, and match a corresponding privacy processing strategy for each sensitivity level. The privacy processing module is used to perform differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy, and generate an intermediate image of the candidate privacy object; The image fusion module is used to fuse the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object.

[0005] In a preferred embodiment, when the image parsing module performs preliminary parsing of the image to be processed to obtain candidate privacy objects of the image to be processed, it is specifically used for: Obtain the image semantic segmentation result of the image to be processed, and identify the target object in the image to be processed based on the image semantic segmentation result; Perform contour detection on the target object to obtain the contour boundary information of the target object; The target objects corresponding to the contour boundary information are filtered by type, and the target objects whose type belongs to the privacy object type set are determined as candidate privacy objects of the image to be processed.

[0006] In a preferred embodiment, when the feature matching module performs feature extraction on the candidate privacy object, generates an object feature vector of the candidate privacy object, and matches the object feature vector with a preset privacy feature library to obtain the matching confidence of the candidate privacy object, it is specifically used for: Perform key point detection on the candidate privacy object and extract the key point location information of the candidate privacy object; Based on the key point location information, a local texture descriptor of the candidate privacy object is generated, and the local texture descriptor is combined into an object feature vector of the candidate privacy object; The object feature vector is compared with the reference feature vector stored in the preset privacy feature library to obtain the similarity between the object feature vector and the reference feature vector; The similarity is quantitatively evaluated with the similarity threshold of the candidate privacy object to obtain the matching confidence of the candidate privacy object.

[0007] In a preferred embodiment, when the mask generation module generates a privacy region mask for the candidate privacy object based on the object feature vector when the privacy attribute of the candidate privacy object is determined to be privacy-related, it is specifically used for: Obtain the privacy attribute judgment result carried in the feature vector of the object; When the privacy attribute judgment result indicates that the candidate privacy object belongs to the privacy category, the spatial distribution information of the candidate privacy object is parsed from the object feature vector; Based on the spatial distribution information, locate the pixel region occupied by the candidate privacy object in the image coordinate system of the image to be processed; The pixel region is masked to obtain the privacy region mask of the candidate privacy object.

[0008] In a preferred embodiment, when the mask generation module performs the judgment result of obtaining the privacy attribute carried in the object feature vector, it is specifically used for: Data structure parsing is performed on the object feature vector to extract feature fields carrying privacy category information from the object feature vector; The feature field is compared with the privacy category feature template recorded in the privacy category mapping table of the candidate privacy object to determine the privacy category matched by the feature field; When the feature field matches any privacy category feature template in the privacy category mapping table, the privacy attribute judgment result is determined as privacy; When the feature field does not match any of the privacy category feature templates in the privacy category mapping table, the privacy attribute judgment result is determined to be non-privacy.

[0009] In a preferred embodiment, when the level division module performs the task of dividing the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence level, and matching a corresponding privacy processing strategy for each sensitivity level, it is specifically used for: The matching confidence level is compared with a preset confidence threshold to determine the confidence interval to which the matching confidence level belongs; Based on the mapping relationship between the confidence interval and the sensitivity level, the sensitivity level of the privacy region corresponding to the privacy region mask is determined; Obtain the correspondence between the sensitivity level and the privacy processing strategy, and select a privacy processing strategy that matches the sensitivity level from the privacy processing strategy set according to the correspondence; The selected privacy processing strategy is determined as the privacy processing strategy corresponding to the privacy region mask.

[0010] In a preferred embodiment, when the level classification module selects a privacy processing policy matching the sensitivity level from the privacy processing policy set according to the correspondence, it is specifically used for: Collect the mapping relationship between the sensitivity level and the privacy processing strategy; Based on the mapping relationship, the sensitivity level is semantically mapped and transformed to obtain the policy feature descriptor corresponding to the sensitivity level; Using the policy feature descriptor as the retrieval basis, feature matching is performed on the privacy processing policy set to obtain the privacy processing policy that matches the policy feature descriptor.

[0011] In a preferred embodiment, when the privacy processing module performs differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy to generate an intermediate image of the candidate privacy object, it is specifically used for: Extract the set of pixel coordinates corresponding to the mask identifier value in the privacy region mask. The set of pixel coordinates is used to indicate the pixel position occupied by the privacy region in the image to be processed. The privacy processing operation type of the candidate privacy object is parsed from the privacy processing strategy; Based on the privacy processing operation type, the pixel positions indicated by the pixel coordinate set are reconstructed and filled to obtain the processed privacy region image block of the candidate privacy object; The pixel positions indicated by the set of pixel coordinates are replaced with the processed privacy region image blocks to generate an intermediate image of the candidate privacy object.

[0012] In a preferred embodiment, when the privacy processing module performs content reconstruction and filling on the pixel positions indicated by the pixel coordinate set according to the privacy processing operation type to obtain the processed privacy region image block of the candidate privacy object, it is specifically used for: When the privacy processing operation type is indicated as content reconstruction and filling, the image content reconstruction process for the pixel positions indicated by the pixel coordinate set is triggered; The spatial topology of the pixel coordinate set in the image to be processed is analyzed, and the neighborhood association relationship of the pixel coordinate set is determined based on the spatial topology. Based on the neighborhood association relationship, texture primitive features that are not affected by privacy processing are extracted around the set of pixel coordinates in the image to be processed. The texture primitive features are used to characterize the visual structure pattern of the neighborhood image region. Using the texture primitive features as the basis for reconstruction, a feature transfer operation is performed on the pixel positions indicated by the pixel coordinate set to generate reconstructed pixel content that matches the visual features in the neighboring image region. The reconstructed pixel content is filled into the pixel positions indicated by the pixel coordinate set to obtain the processed privacy region image block of the candidate privacy object.

[0013] In a preferred embodiment, when the image fusion module performs image fusion of the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object, it is specifically used for: Based on the coverage area of ​​the privacy region mask, the processed privacy region and the unprocessed region are separated from the intermediate image; At the boundary between the processed privacy region and the unprocessed region, extract the boundary pixel sequence of the processed privacy region and the adjacent pixel sequence of the unprocessed region; Using the boundary pixel sequence and the adjacent pixel sequence as references, pixel transition processing is performed on the processed privacy region and the unprocessed region to obtain the transition fusion band of the candidate privacy object; The transition fusion band is stitched together with the processed privacy region and the unprocessed region to obtain the output image of the candidate privacy object.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention accurately locates candidate privacy objects through image analysis and feature matching, generates object feature vectors and completes privacy confidence assessment, and automatically generates privacy region masks by combining spatial distribution information, thereby improving the accuracy and completeness of privacy content recognition.

[0015] 2. This invention classifies sensitivity levels based on matching confidence and matches differentiated processing strategies to reconstruct content and fill pixels in privacy areas. By using boundary transition and image fusion, it achieves a natural connection between the processed area and the original image, thereby improving privacy protection and the visual integrity of the output image. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of an image privacy content recognition and protection system provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

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

[0018] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “said” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0019] Depending on the context, the word "if" or "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0020] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0021] In practice, the server-side equipment deployed in an image privacy content recognition and protection system may consist of one or more devices. This image privacy content recognition and protection system can be implemented as a business instance, a virtual machine, or hardware devices. For example, it can be implemented as a business instance deployed on one or more devices in a cloud node. Simply put, it can be understood as software deployed on a cloud node to provide image privacy content recognition and protection to various user terminals. Alternatively, it can be implemented as a virtual machine deployed on one or more devices in a cloud node, with application software installed to manage various user terminals. Or, it can also be implemented as a server composed of numerous identical or different types of hardware devices, with one or more hardware devices configured to provide image privacy content recognition and protection to various user terminals.

[0022] In terms of implementation, the image privacy content recognition and protection system and the user terminal are mutually compatible. That is, if the image privacy content recognition and protection system is implemented as an application installed on a cloud service platform, then the user terminal is a client that establishes a communication connection with the application; or if the image privacy content recognition and protection system is implemented as a website, then the user terminal is implemented as a webpage; or if the image privacy content recognition and protection system is implemented as a cloud service platform, then the user terminal is implemented as a mini-program in an instant messaging application.

[0023] like Figure 1 The diagram shown is a system architecture diagram of an image privacy content recognition and protection system provided in an embodiment of the present invention.

[0024] The image privacy content recognition and protection system described in this invention can be set up in a cloud server. In terms of implementation, it can be implemented as one or more service devices, or as an application installed in the cloud (e.g., a mobile service operator's server, server cluster, etc.), or it can be developed into a website. Depending on the functions implemented, the image privacy content recognition and protection system may include an image parsing module, a feature matching module, a mask generation module, a level classification module, a privacy processing module, and an image fusion module. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0025] In this embodiment of the invention, in an image privacy content recognition and protection system, each of the above-mentioned modules can be implemented independently and can call other modules. Here, "calling" can be understood as a module connecting to multiple modules of another type and providing corresponding services to those connected modules. In the image privacy content recognition and protection system provided by this embodiment of the invention, without modifying the program code, the applicability of the system architecture can be adjusted by adding modules and directly calling them, achieving cluster-based horizontal expansion to quickly and flexibly expand the image privacy content recognition and protection system. In practical applications, the above modules can be set in the same device or different devices, or they can be set in a virtual device, such as a service instance in a cloud server.

[0026] The following describes, with reference to specific embodiments, each component and its specific workflow of an image privacy content recognition and protection system: The image parsing module is used to perform preliminary parsing on the image to be processed to obtain candidate privacy objects of the image to be processed; In this embodiment of the invention, when the image parsing module performs preliminary parsing of the image to be processed to obtain candidate privacy objects of the image to be processed, it is specifically used for: Obtain the image semantic segmentation result of the image to be processed, and identify the target object in the image to be processed based on the image semantic segmentation result; Perform contour detection on the target object to obtain the contour boundary information of the target object; The target objects corresponding to the contour boundary information are filtered by type, and the target objects whose type belongs to the privacy object type set are determined as candidate privacy objects of the image to be processed.

[0027] The image to be processed is retrieved. This image is either a raw image captured by an image acquisition device or an image file stored in the system. The image to be processed is then processed using a preset image semantic segmentation method. This method can divide the image according to different semantic categories. During the segmentation process, the image content is analyzed pixel by pixel to distinguish different entity regions in the image, thus obtaining the image semantic segmentation result of the image to be processed. Based on the semantic segmentation result, all entities contained in the image are identified one by one. These entities are uniformly identified as target objects in the image to be processed, ensuring that all target objects can be completely identified without omissions or misidentifications. The semantic segmentation method is based on preset image semantic annotation rules, which can accurately distinguish different types of image entities.

[0028] For each identified target object, contour detection is performed. Edge detection is used to identify the edge pixels of the target object one by one, capturing the pixel changes of the target object's edge. Continuous edge pixels are connected to form a complete edge contour of the target object. Then, the contour is denoised to remove redundant and interfering pixels, ensuring the integrity and accuracy of the contour. Finally, the contour boundary information corresponding to each target object is obtained. This information can clearly define the specific location and shape range of the target object in the image. The edge detection method used in the detection process determines the edge position by identifying abrupt changes in pixel gray values ​​in the image, ensuring that the contour boundary information completely matches the actual shape of the target object.

[0029] A pre-established set of privacy object types is created, which is derived from the system's preset privacy protection specifications. This set includes all object types that require privacy protection, covering various privacy objects such as those related to personal identity and private parts. For each target object corresponding to the contour boundary information, the type is determined by comparing the characteristics of the target object with the characteristics of each category in the privacy object type set. If the type of the target object completely matches a category in the privacy object type set, the target object is identified as a candidate privacy object for the image to be processed. If it does not match, it is excluded, ensuring the accuracy of the candidate privacy objects. The screening process strictly follows the classification criteria of the privacy object type set, and there are no cases of misjudgment or omission.

[0030] The beneficial effects are that the target object can be directly identified through the image semantic segmentation results, the image content structured parsing can be completed quickly, the whole domain entity extraction can be achieved without omission, and a complete and reliable basic target set can be provided for subsequent privacy judgment.

[0031] By performing contour detection on the target object to obtain boundary information, the object's position and shape can be accurately located, background noise interference can be eliminated, the positioning accuracy of privacy areas can be improved, and false positives and false negatives caused by blurred boundaries can be avoided.

[0032] By filtering target objects based on a set of privacy object types, non-privacy targets can be quickly eliminated, reducing the computational load of subsequent modules, ensuring the compliance and accuracy of candidate privacy objects from the source, and improving the overall identification efficiency.

[0033] It adopts a three-level progressive analysis of semantic segmentation, contour detection and type filtering, which takes into account both the comprehensiveness of recognition and the accuracy of filtering, effectively solving the problems of inaccuracy and low efficiency of traditional recognition, and laying a solid front-end foundation for privacy protection throughout the entire process.

[0034] The feature matching module is used to extract features from the candidate privacy object, generate an object feature vector of the candidate privacy object, and match the object feature vector with a preset privacy feature library to obtain the matching confidence of the candidate privacy object. In this embodiment of the invention, when the feature matching module performs feature extraction on the candidate privacy object, generates an object feature vector of the candidate privacy object, and matches the object feature vector with a preset privacy feature library to obtain the matching confidence of the candidate privacy object, it is specifically used for: Perform key point detection on the candidate privacy object and extract the key point location information of the candidate privacy object; Based on the key point location information, a local texture descriptor of the candidate privacy object is generated, and the local texture descriptor is combined into an object feature vector of the candidate privacy object; The object feature vector is compared with the reference feature vector stored in the preset privacy feature library to obtain the similarity between the object feature vector and the reference feature vector; The similarity is quantitatively evaluated with the similarity threshold of the candidate privacy object to obtain the matching confidence of the candidate privacy object.

[0035] The candidate privacy objects output by the image analysis module are scanned region by region. The distinctive feature points on the object are identified by pixel grayscale changes and contour inflection point localization. The specific coordinate information of each feature point in the image is recorded one by one to form the complete key point location information of the candidate privacy object. This information comes from the actual image pixel distribution of the candidate privacy object and can accurately reflect the core structural features of the object.

[0036] Based on the extracted key point location information, a local feature region is defined with each key point as the center. Texture features are extracted from the pixel arrangement and brightness variation patterns within the region to form a local texture descriptor for the corresponding key point. All local texture descriptors corresponding to the key points are integrated and arranged in a fixed order to form an object feature vector that can fully represent the overall features of the candidate privacy object. This vector fully preserves the texture and structural feature information of the object.

[0037] The system retrieves a pre-built and stored privacy feature library containing reference feature vectors corresponding to various standard privacy objects. The generated object feature vector is compared with each reference feature vector in the library dimension by dimension. The similarity of features in each dimension is compared one by one, and the number of overlaps and the degree of similarity in feature dimensions are counted. Finally, the similarity value between the object feature vector and the corresponding reference feature vector is obtained. The similarity value directly reflects the degree of feature similarity between the candidate privacy object and the standard privacy object.

[0038] The calculated similarity values ​​are compared item by item with the system's pre-set similarity threshold to determine whether the similarity value reaches or exceeds the threshold. At the same time, the difference between the similarity value and the threshold is combined for a comprehensive evaluation. The smaller the difference, the higher the evaluation result, and the larger the difference, the lower the evaluation result. Through this quantitative evaluation method, the matching confidence of the candidate privacy object is finally obtained. This confidence is used to measure the degree to which the object is credible as a real privacy object.

[0039] The beneficial effect is that by performing key point detection and extracting location information on candidate privacy objects, the core feature points of the objects can be accurately located, the structural recognizability is fully preserved, and a stable and reliable localization foundation is provided for subsequent texture extraction and feature matching.

[0040] By generating local texture descriptors based on key point location information and combining them into object feature vectors, the visual details and texture patterns of candidate privacy objects can be comprehensively characterized, improving the completeness and discriminativeness of feature expression.

[0041] By comparing the object's feature vector with the reference feature vector in the preset privacy feature library, the degree of feature similarity can be quantified, and an objective similarity value can be obtained, providing a quantifiable core basis for privacy attribute determination.

[0042] By quantifying similarity and similarity thresholds, a stable and reliable matching confidence level can be output, enabling standardized determination of privacy objects, improving recognition consistency and reducing the false positive rate.

[0043] By employing a four-level process of key point detection, texture description, vector comparison, and confidence assessment, the accuracy of privacy object feature representation and matching is significantly improved, effectively compensating for the feature loss and ambiguity problems of traditional recognition methods.

[0044] The mask generation module is used to generate a privacy region mask for the candidate privacy object based on the object feature vector when the privacy attribute of the candidate privacy object is determined to be private. In this embodiment of the invention, when the mask generation module generates a privacy region mask for the candidate privacy object based on the object feature vector when the privacy attribute of the candidate privacy object is determined to be private, it is specifically used for: Obtain the privacy attribute judgment result carried in the feature vector of the object; When the privacy attribute judgment result indicates that the candidate privacy object belongs to the privacy category, the spatial distribution information of the candidate privacy object is parsed from the object feature vector; Based on the spatial distribution information, locate the pixel region occupied by the candidate privacy object in the image coordinate system of the image to be processed; The pixel region is masked to obtain the privacy region mask of the candidate privacy object.

[0045] When the mask generation module performs the judgment result of obtaining the privacy attribute carried in the feature vector of the object, it is specifically used for: Data structure parsing is performed on the object feature vector to extract feature fields carrying privacy category information from the object feature vector; The feature field is compared with the privacy category feature template recorded in the privacy category mapping table of the candidate privacy object to determine the privacy category matched by the feature field; When the feature field matches any privacy category feature template in the privacy category mapping table, the privacy attribute judgment result is determined as privacy; When the feature field does not match any of the privacy category feature templates in the privacy category mapping table, the privacy attribute judgment result is determined to be non-privacy.

[0046] The system retrieves candidate privacy object-related data output by the feature matching module, focusing on extracting the privacy attribute judgment results carried in the object feature vector. These results are clear conclusions drawn by the feature matching module after matching the object feature vector with a preset privacy feature library. They are directly used to determine whether a candidate privacy object belongs to a privacy category, ensuring that the obtained privacy attribute judgment results correspond one-to-one with the candidate privacy objects, without confusion or mismatch. All data sources are valid results from previous feature extraction and matching, ensuring the accuracy and reliability of the judgment basis.

[0047] When the obtained privacy attribute judgment result clearly indicates that the candidate privacy object belongs to the privacy category, the object feature vector corresponding to the candidate privacy object is fully analyzed. Spatial distribution information that represents the specific location and morphological range of the candidate privacy object in the image is extracted from the object feature vector. During the analysis process, each dimension of the object feature vector is filtered one by one to accurately separate the feature information related to the spatial location. This ensures that the analyzed spatial distribution information can fully reflect the specific distribution of the candidate privacy object in the image to be processed and is consistent with the original features of the object feature vector.

[0048] Based on the parsed spatial distribution information and combined with the image coordinate system parameters of the image to be processed, the specific position of the candidate privacy object in the image is located pixel by pixel. First, the edge pixel coordinates of the candidate privacy object are determined, and then the complete outline of the candidate privacy object is delineated according to the edge pixel coordinates. In this way, the entire pixel area occupied by the candidate privacy object in the image coordinate system is determined. This pixel area accurately corresponds to the actual display range of the candidate privacy object in the image to be processed, without any omissions or extra pixels.

[0049] The pixel regions occupied by the identified candidate privacy objects are uniformly masked. During the labeling process, each pixel is marked to clearly distinguish it from other non-privacy pixel regions in the image to be processed. The marked pixel regions are then assigned a uniform mask identifier to form a complete privacy region mask. This privacy region mask can accurately cover the entire pixel range of the candidate privacy object, clearly define the specific location and range of the privacy region, and provide a clear identification basis for subsequent privacy region hiding or processing. This ensures the accuracy and completeness of the mask generation and complies with the technical specifications for image privacy protection.

[0050] The object feature vector is subjected to complete data structure parsing. According to the storage format of the object feature vector, the information contained therein is decomposed field by field. The focus is on locating the feature field carrying the privacy attribute judgment result. This feature field is generated by the feature matching module after comparing the features of the candidate privacy objects in the early stage. This ensures that the relevant data carrying the privacy attribute judgment result is accurately extracted from the object feature vector. No information related to the privacy attribute judgment is missed in the parsing process, ensuring that the extracted privacy attribute judgment result is complete and accurate. The data sources are all valid results after the previous feature extraction and matching, which meet the classification requirements.

[0051] The extracted feature fields carrying privacy category information are comprehensively compared with a pre-set privacy category mapping table. The privacy category mapping table is a standard template based on common privacy object features, containing feature templates corresponding to various privacy objects. During the comparison process, the details of each feature field are checked one by one to confirm the degree of matching between the feature field and the privacy category feature template, and to clarify the privacy category corresponding to the feature field. This ensures that the matching process is rigorous and that there is no category confusion. The data source is the relevant results after the previous feature extraction and matching to ensure the accuracy of the comparison results.

[0052] When a feature field completely matches any privacy category feature template in the privacy category mapping table, the privacy attribute judgment result is directly determined as privacy, indicating that the candidate privacy object belongs to the category that needs privacy protection. Subsequently, the object will be processed in accordance with relevant privacy protection specifications to ensure that privacy information is not leaked. This judgment result is directly related to the generation of the subsequent privacy region mask. The data source is the clear result after feature comparison, which meets the classification requirements.

[0053] When the feature field does not match any of the privacy category feature templates in the privacy category mapping table, the privacy attribute judgment result is directly determined as non-privacy, indicating that the candidate privacy object does not belong to the category that requires key privacy protection. There is no need to perform privacy masking processing on it in the future. This judgment result is also based on the effective data of the previous feature comparison, ensuring the accuracy and reliability of the judgment result and complying with the relevant classification technical specifications.

[0054] The beneficial effect is that by directly obtaining the privacy attribute judgment results carried in the object feature vector, the privacy attribute confirmation can be completed quickly without repeating feature calculations, effectively improving the efficiency of privacy judgment and the consistency of the process.

[0055] After confirming that the candidate privacy object belongs to the privacy category, the spatial distribution information is analyzed from the object feature vector, which can accurately obtain the object's position and shape in the image, ensuring that the positioning basis is reliable.

[0056] By locating pixel regions in the image coordinate system based on spatial distribution information, the actual coverage area of ​​candidate privacy objects can be accurately delineated, avoiding region offset or omission and improving mask positioning accuracy.

[0057] Applying a uniform mask to pixel regions can clearly distinguish between privacy and non-privacy regions, forming a standardized privacy region mask that provides an accurate regional basis for subsequent differentiated processing.

[0058] This step, through a coherent process of attribute determination, spatial parsing, region positioning, and mask marking, enables the automatic and accurate generation of privacy regions, significantly improving the integrity of identification and the reliability of protection.

[0059] By parsing the data structure of the object feature vector and extracting the privacy category feature field, we can accurately locate privacy attribute-related information, avoid interference from irrelevant data, and provide a clear and reliable source of information for subsequent privacy category determination.

[0060] By comparing feature fields with the privacy category mapping table template, the corresponding privacy category can be quickly matched, improving the efficiency of judgment, ensuring that the privacy attribute identification process is standardized and consistent, and reducing human judgment errors.

[0061] If a feature field matches any template, it is determined to be private, which can quickly identify the object that needs protection, promptly initiate the privacy protection process, and improve the response speed and timeliness of privacy content protection.

[0062] If a feature field does not match any template, it is determined to be non-privacy. This can effectively exclude non-privacy objects, reduce unnecessary processing overhead, optimize system resource allocation, and improve overall operating efficiency and accuracy.

[0063] The level division module is used to divide the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence, and match a corresponding privacy processing strategy for each sensitivity level. In this embodiment of the invention, when the level division module performs the operation of dividing the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence level, and matching a corresponding privacy processing strategy for each sensitivity level, it is specifically used for: The matching confidence level is compared with a preset confidence threshold to determine the confidence interval to which the matching confidence level belongs; Based on the mapping relationship between the confidence interval and the sensitivity level, the sensitivity level of the privacy region corresponding to the privacy region mask is determined; Obtain the correspondence between the sensitivity level and the privacy processing strategy, and select a privacy processing strategy that matches the sensitivity level from the privacy processing strategy set according to the correspondence; The selected privacy processing strategy is determined as the privacy processing strategy corresponding to the privacy region mask.

[0064] When the level classification module selects a privacy processing strategy matching the sensitivity level from the privacy processing strategy set according to the correspondence, it is specifically used for: Collect the mapping relationship between the sensitivity level and the privacy processing strategy; Based on the mapping relationship, the sensitivity level is semantically mapped and transformed to obtain the policy feature descriptor corresponding to the sensitivity level; Using the policy feature descriptor as the retrieval basis, feature matching is performed on the privacy processing policy set to obtain the privacy processing policy that matches the policy feature descriptor.

[0065] The system retrieves the previously generated matching confidence scores, which are derived from the matching evaluation results of the feature matching module on candidate privacy objects. These scores are the core basis for determining the sensitivity level of privacy regions. Simultaneously, the system retrieves the preset confidence thresholds and compares each matching confidence score with the preset threshold. During the comparison process, the preset interval division standard is strictly followed to clarify the specific confidence interval in which the current matching confidence score belongs. This ensures that each matching confidence score corresponds to a unique confidence interval. The comparison process does not add any extra calculations; the interval assignment is determined solely through direct comparison, ensuring the accuracy and uniqueness of the interval judgment and providing a foundation for subsequent sensitivity level division.

[0066] Based on the preset mapping relationship between confidence intervals and sensitivity levels, which is pre-defined in conjunction with privacy protection needs and actual application scenarios, the specific sensitivity level corresponding to each confidence interval is clearly defined. After determining the confidence interval to which the matching confidence level belongs, the corresponding sensitivity level is directly matched to ensure that the division of sensitivity levels is highly consistent with the matching confidence level, accurately reflecting the sensitivity of the privacy area. This sensitivity level directly corresponds to the protection priority of the privacy area, providing a clear basis for the matching of subsequent privacy processing strategies. The data sources are all from the previous matching confidence results and preset mapping rules, ensuring the rationality of the level division.

[0067] The system retrieves a pre-defined table of sensitivity levels and privacy processing strategies. This table is based on the privacy protection requirements of different sensitivity levels, clearly defining the specific processing requirements and operating procedures for each sensitivity level. It covers privacy protection methods in different scenarios. Based on the determined sensitivity level of the privacy area, the system accurately searches the table and selects a privacy processing strategy that perfectly matches the sensitivity level from the pre-defined set of privacy processing strategies. This ensures that the selected strategy can meet the protection requirements of the privacy area at that sensitivity level and fits the actual scenario of privacy processing.

[0068] The selected privacy processing strategy, matching the sensitivity level, is directly designated as the privacy processing strategy corresponding to the privacy area mask. The specific implementation requirements and operational procedures of this privacy processing strategy are clearly defined to ensure its accurate application to the corresponding privacy area. This achieves targeted protection for privacy areas with different sensitivity levels, ensuring the security of privacy information while avoiding resource waste caused by over-protection. A one-to-one correspondence is established between the privacy processing strategy, the privacy area mask, and the sensitivity level. The data sources are all from previous sensitivity level assessments and pre-defined strategy correspondences, complying with relevant data security protection standards and ensuring the standardization and effectiveness of privacy processing.

[0069] The system retrieves the pre-defined mapping relationship between sensitivity levels and privacy processing policies. This mapping relationship is pre-set based on privacy protection needs and actual application scenarios, covering privacy processing policies corresponding to all sensitivity levels. During the data collection process, the correspondence between each sensitivity level and its corresponding privacy processing policy is reviewed one by one to ensure that no sensitivity level mapping information is missed. At the same time, the accuracy of the mapping relationship is verified to avoid mapping errors, ensuring that the collected mapping relationship can be directly used for subsequent policy matching. All mapping relationships come from the system's pre-defined privacy protection specifications, conform to the actual needs of privacy processing, and comply with relevant data security classification requirements.

[0070] Based on the collected mapping relationship between sensitivity levels and privacy processing policies, for each currently determined sensitivity level, the sensitivity levels in the mapping relationship are compared one by one to find the mapping entry that is completely consistent with the current sensitivity level. Then, according to the policy description corresponding to the mapping entry, the current sensitivity level is semantically mapped and transformed, converting the feature information of the sensitivity level into the corresponding policy feature descriptor. The transformation process strictly follows the requirements of the mapping relationship to ensure that the policy feature descriptor can accurately reflect the privacy processing requirements corresponding to the sensitivity level, remain consistent with the core features of the sensitivity level, and have no deviation, thus ensuring the accuracy of subsequent policy matching.

[0071] Using the generated policy feature descriptors as the sole retrieval criterion, each policy in the privacy processing policy set is compared one by one. During the comparison process, the core features of each policy are the focus, and it is determined whether the features of each policy completely match the policy feature descriptors. The core content of each policy, such as its applicable scenarios, processing methods, and protection requirements, is checked one by one during the comparison. Privacy processing policies that do not match the policy feature descriptors are excluded, and privacy processing policies that are highly consistent with the policy feature descriptors and can meet the current sensitivity level privacy protection requirements are selected. This ensures that the selected policies are fully adapted to the sensitivity level, meet the actual needs of privacy processing, comply with relevant data security protection specifications, and provide a clear execution basis for subsequent privacy processing.

[0072] The beneficial effect is that by comparing the matching confidence level with the confidence threshold to determine the range, the risk level of privacy areas can be quantitatively distinguished, making the classification of sensitivity levels based on evidence and avoiding the problem of unbalanced protection caused by subjective judgment.

[0073] By classifying privacy risks based on the mapping relationship between confidence intervals and sensitivity levels, a standardized stratification of privacy risks can be achieved, enabling precise matching of the protection strength and risk level of different privacy areas and improving the rationality of protection.

[0074] By obtaining the correspondence between sensitivity levels and policies and accurately selecting appropriate policies, privacy protection methods can be highly aligned with regional risks, achieving the optimal configuration of protection strength and processing methods.

[0075] By defining the selected strategy as a privacy-specific processing strategy, precise protection can be achieved with a one-strategy-per-zone approach, avoiding over-protection or under-protection by general strategies and improving overall protection efficiency.

[0076] By establishing a mapping relationship between sensitivity levels and privacy processing policies, a stable correspondence rule between levels and policies can be established, providing a unified standard for subsequent policy matching and improving the standardization and consistency of policy selection.

[0077] By performing semantic mapping transformation on sensitivity levels based on mapping relationships, abstract levels can be transformed into concrete strategy feature descriptors, making strategy retrieval more targeted and improving matching accuracy and efficiency.

[0078] Using policy feature descriptors as the basis for feature matching can accurately select suitable privacy processing policies, avoid policy mismatch, and achieve a precise match between protection strength and sensitivity level.

[0079] The privacy processing module is used to perform differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy, and generate an intermediate image of the candidate privacy object; In this embodiment of the invention, when the privacy processing module performs differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy to generate an intermediate image of the candidate privacy object, it is specifically used for: Extract the set of pixel coordinates corresponding to the mask identifier value in the privacy region mask. The set of pixel coordinates is used to indicate the pixel position occupied by the privacy region in the image to be processed. The privacy processing operation type of the candidate privacy object is parsed from the privacy processing strategy; Based on the privacy processing operation type, the pixel positions indicated by the pixel coordinate set are reconstructed and filled to obtain the processed privacy region image block of the candidate privacy object; The pixel positions indicated by the set of pixel coordinates are replaced with the processed privacy region image blocks to generate an intermediate image of the candidate privacy object.

[0080] When the privacy processing module performs content reconstruction and filling on the pixel positions indicated by the pixel coordinate set according to the privacy processing operation type to obtain the processed privacy region image block of the candidate privacy object, it is specifically used for: When the privacy processing operation type is indicated as content reconstruction and filling, the image content reconstruction process for the pixel positions indicated by the pixel coordinate set is triggered; The spatial topology of the pixel coordinate set in the image to be processed is analyzed, and the neighborhood association relationship of the pixel coordinate set is determined based on the spatial topology. Based on the neighborhood association relationship, texture primitive features that are not affected by privacy processing are extracted around the set of pixel coordinates in the image to be processed. The texture primitive features are used to characterize the visual structure pattern of the neighborhood image region. Using the texture primitive features as the basis for reconstruction, a feature transfer operation is performed on the pixel positions indicated by the pixel coordinate set to generate reconstructed pixel content that matches the visual features in the neighboring image region. The reconstructed pixel content is filled into the pixel positions indicated by the pixel coordinate set to obtain the processed privacy region image block of the candidate privacy object.

[0081] The previously generated privacy region mask is retrieved. This privacy region mask originates from the image region marking of candidate privacy objects. All pixel coordinates corresponding to the mask identifier value are clearly defined. These pixel coordinates are collected one by one and organized into a pixel coordinate set. Each pixel coordinate corresponds to a pixel in the image to be processed that belongs to the privacy region. By using the pixel coordinate set, the specific location of the privacy region in the image to be processed is accurately located, ensuring that the pixels of each privacy region are accurately recorded without omissions or misselections. All pixel coordinates come from the actual pixel distribution of the image to be processed, closely matching the actual range of the privacy region, providing accurate location information for subsequent privacy processing.

[0082] From the previously determined privacy handling strategy, each part of the strategy was broken down and analyzed to extract the privacy handling operation types explicitly specified for candidate privacy objects. The handling methods corresponding to different privacy areas were clearly distinguished to ensure that the extracted handling operation types matched the privacy attributes and sensitivity of the candidate privacy objects, fully complied with the relevant requirements of the privacy handling strategy, and did not add any additional handling methods. All handling operation types came from the preset privacy handling strategy, which is in line with the actual needs of privacy protection and provides a clear operational basis for subsequent privacy area handling.

[0083] Based on the parsed privacy processing operation type, content reconstruction and filling are performed on each pixel position indicated by the pixel coordinate set. During the reconstruction and filling process, the requirements of the privacy processing operation type are strictly followed. Combined with the overall tone and texture features of the image to be processed, the filling content is generated to be consistent with the surrounding image environment. This ensures that the filled pixels have the same image style as the surrounding non-privacy areas and do not have obvious abruptness. The filling process is performed pixel by pixel to ensure that the pixels of each privacy area are accurately reconstructed, generating a processed privacy area image block. This image block is completely consistent with the pixel range of the original privacy area and can completely cover the original privacy area.

[0084] The processed privacy region image block is precisely aligned with the image to be processed. According to the position indicated by the pixel coordinate set, the corresponding original privacy region pixels in the image to be processed are replaced one by one. During the replacement process, each pixel coordinate is strictly matched to ensure that the processed privacy region image block accurately covers the original privacy region without misalignment, omission, or redundant replacement. After the replacement is completed, the processed privacy region is integrated with the non-privacy region of the original image to form a complete intermediate image of the candidate privacy object. This intermediate image retains the non-privacy part of the original image and performs standardized processing on the privacy region, which meets the relevant privacy protection requirements. The data sources are all previously generated privacy region masks, processed privacy region image blocks, and the original image to be processed, ensuring the integrity and accuracy of the intermediate image.

[0085] When the privacy processing operation type is explicitly indicated as content reconstruction and filling, the image content reconstruction process for the pixel position indicated by the pixel coordinate set is immediately triggered. This process strictly follows the preset privacy processing specifications, uses the pixel coordinate set as the core basis, accurately locates the pixel area that needs to be reconstructed, and ensures that the reconstruction process only applies to the privacy area and does not affect the non-privacy area in the image. All operations revolve around the privacy area of ​​the candidate privacy object, which fits the core requirements of content reconstruction and filling and ensures the targeting and accuracy of the reconstruction process.

[0086] This method analyzes the spatial topology of the pixel coordinate set in the image to be processed, systematically examines the specific location of each pixel coordinate in the image, clarifies the adjacency relationships, distribution patterns, and overall arrangement of each pixel, clearly understands the image area range and structural characteristics covered by the pixel coordinate set, and determines the neighborhood association relationship between each pixel in the pixel coordinate set and its surrounding pixels based on this information. It identifies which pixels belong to adjacent regions and which pixels constitute complete privacy region blocks, providing accurate spatial basis for subsequent texture feature extraction and content reconstruction.

[0087] Based on the established neighborhood relationships, the non-privacy region surrounding the set of pixel coordinates in the image to be processed is accurately located. This region is not affected by any privacy processing and retains the original image texture and visual features. By analyzing each pixel, representative texture primitive features within this region are extracted. These texture primitive features can fully characterize the visual structure pattern of the neighborhood image region, including core information such as pixel hue, texture direction, and brightness distribution. This ensures that the extracted features can truly reflect the visual characteristics of the neighborhood image and provide a reliable basis for subsequent content reconstruction.

[0088] Using the extracted texture primitive features as the core basis for content reconstruction, a feature transfer operation is performed on each pixel position indicated by the set of pixel coordinates. Specifically, the texture primitive features of the non-privacy neighborhood are transferred to the pixel position to be reconstructed according to the spatial topology and visual rules of the set of pixel coordinates. This ensures that the reconstructed pixel content is highly compatible with the visual features of the neighborhood image region and does not appear abrupt. During the transfer process, the texture style of the neighborhood image is strictly followed to ensure the naturalness and rationality of the reconstructed content.

[0089] The generated reconstructed pixel content is filled pixel by pixel into all pixel positions indicated by the set of pixel coordinates. During the filling process, each pixel coordinate is strictly matched to ensure that each pixel position that needs to be reconstructed is accurately filled without omission, misalignment or duplicate filling. After filling is completed, all reconstructed pixels are integrated to form a complete processed privacy region image block. This image block is consistent with the visual style of the neighboring non-privacy region. It not only realizes the content reconstruction of the privacy region, but also ensures the overall coordination of the image, which complies with the relevant specifications for privacy protection and image processing, and can meet the actual needs of privacy region processing.

[0090] The beneficial effect is that by extracting the set of pixel coordinates corresponding to the privacy region mask, the pixel position of the privacy region in the image can be accurately located, providing an accurate location basis for subsequent privacy processing and avoiding processing range deviation.

[0091] Analyzing privacy processing operation types from privacy processing strategies can clarify differentiated processing methods, match processing operations with sensitivity levels, and improve the targeting and rationality of privacy protection.

[0092] By reconstructing and filling pixel locations according to the type of processing operation, image blocks of privacy regions that adapt to the surrounding visual features can be generated, thus protecting privacy while maintaining the overall harmony of the image.

[0093] By replacing the original pixel positions with processed privacy region image blocks and generating an intermediate image, the normalization of the privacy region can be completed, providing a structurally complete base image for subsequent image fusion.

[0094] By triggering the image content reconstruction process, exclusive content replacement can be performed on privacy areas, protecting privacy information while avoiding interference with non-privacy areas and improving processing accuracy.

[0095] By analyzing the spatial topology of pixel coordinate sets and determining neighborhood relationships, the connection between privacy regions and surrounding images can be accurately located, providing a reliable spatial basis for texture feature extraction.

[0096] Extracting unaffected texture primitive features around privacy regions preserves the visual structural patterns of the original image, ensuring that the reconstructed content remains highly consistent with the overall image style.

[0097] Using texture primitive features for feature transfer operations can make the reconstructed pixel content fit the visual features of the neighborhood, eliminate the abruptness after privacy processing, and improve the overall visual integrity of the image.

[0098] By filling the target pixel position with the reconstructed pixel content and generating the processed image block, the privacy area can be completely covered, achieving the dual effect of privacy protection and natural image presentation.

[0099] The image fusion module is used to fuse the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object.

[0100] In this embodiment of the invention, when the image fusion module performs image fusion of the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object, it is specifically used for: Based on the coverage area of ​​the privacy region mask, the processed privacy region and the unprocessed region are separated from the intermediate image; At the boundary between the processed privacy region and the unprocessed region, extract the boundary pixel sequence of the processed privacy region and the adjacent pixel sequence of the unprocessed region; Using the boundary pixel sequence and the adjacent pixel sequence as references, pixel transition processing is performed on the processed privacy region and the unprocessed region to obtain the transition fusion band of the candidate privacy object; The transition fusion band is stitched together with the processed privacy region and the unprocessed region to obtain the output image of the candidate privacy object.

[0101] The previously generated privacy region mask is retrieved. This mask clearly identifies the pixel regions in the image to be processed that require privacy protection. Based on the coverage of this privacy region mask, each region in the image to be processed is checked pixel by pixel. This accurately distinguishes between the processed privacy regions and the unprocessed regions that retain the original image features. This ensures that the boundaries between the two regions are clearly distinguishable and that there is no overlap or omission. All region divisions are based on the clear identification of the privacy region mask, which meets the core requirements of privacy protection, complies with relevant image processing specifications, and ensures the accuracy and relevance of subsequent fusion operations.

[0102] For the separated processed and unprocessed privacy regions, the boundary between them is accurately located, and each pixel position at the boundary is identified one by one. The pixel sequence of the edge of the processed privacy region is extracted. This sequence completely covers the boundary range of the processed region. At the same time, the pixel sequence adjacent to the boundary in the unprocessed region is extracted to ensure that these two pixel sequences can accurately correspond to all pixel positions at the boundary, clearly presenting the connection relationship between the two regions. This provides an accurate pixel basis for subsequent transition fusion. All pixel sequences are from the actual pixel distribution of the image to be processed, ensuring the integrity and accuracy of the sequence.

[0103] Using the extracted pixel sequences of the processed privacy region boundary and the adjacent pixel sequences of the unprocessed region as core references, a transition processing is performed on the pixels at the junction of the two regions. Specifically, the pixel features of the processed privacy region are naturally connected with the pixel features of the unprocessed region. The hue, texture, and brightness of the pixels at the junction are adjusted to maintain a consistent visual style between the two regions, eliminating the abruptness at the junction and forming a smooth transition fusion region. This transition fusion zone can perfectly connect the processed privacy region and the unprocessed region, ensuring the overall harmony and naturalness of the image, and meeting the relevant technical requirements of image processing.

[0104] The generated transition fusion band is precisely stitched together with the processed privacy area and the unprocessed area. During the stitching process, the spatial relationship between each area is strictly followed to ensure that the processed privacy area, the transition fusion band, and the unprocessed area can be seamlessly connected without pixel misalignment, overlap, or omission. After stitching, a complete candidate privacy object output image is formed. This output image not only retains the original image features of the unprocessed area, but also protects the sensitive area through privacy processing. At the same time, the transition fusion band achieves a natural connection between each area, ensuring the integrity and visual consistency of the output image, conforming to the relevant specifications of image fusion, and meeting the dual requirements of privacy protection and image presentation.

[0105] The beneficial effect is that by separating the processed and unprocessed areas based on the coverage of the privacy region mask, the image processing boundaries can be accurately defined, avoiding region confusion and providing a clear processing foundation for subsequent smooth fusion.

[0106] Extracting the boundary pixel sequence and adjacent pixel sequence at the region boundary can completely obtain the visual feature information of the junction, providing a real and reliable reference for pixel transition processing.

[0107] By using two types of pixel sequences as references for pixel transition processing, the visual disconnect between the processed area and the original area can be eliminated, generating a natural and coherent transition fusion zone and improving the overall image quality.

[0108] By combining the transition blending zone with the two types of regions, a complete and unified output image can be formed, which maintains the integrity of the image structure and visual consistency while ensuring privacy protection.

[0109] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0110] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An image privacy content recognition and protection system, characterized in that, The system includes an image parsing module, a feature matching module, a mask generation module, a hierarchy classification module, a privacy processing module, and an image fusion module, wherein: The image parsing module is used to perform preliminary parsing on the image to be processed to obtain candidate privacy objects of the image to be processed; The feature matching module is used to extract features from the candidate privacy object, generate an object feature vector of the candidate privacy object, and match the object feature vector with a preset privacy feature library to obtain the matching confidence of the candidate privacy object. The mask generation module is used to generate a privacy region mask for the candidate privacy object based on the object feature vector when the privacy attribute of the candidate privacy object is determined to be private. The level division module is used to divide the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence, and match a corresponding privacy processing strategy for each sensitivity level. The privacy processing module is used to perform differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy, and generate an intermediate image of the candidate privacy object; The image fusion module is used to fuse the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object.

2. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the image parsing module performs preliminary parsing of the image to be processed to obtain candidate privacy objects, it is specifically used for: Obtain the image semantic segmentation result of the image to be processed, and identify the target object in the image to be processed based on the image semantic segmentation result; Perform contour detection on the target object to obtain the contour boundary information of the target object; The target objects corresponding to the contour boundary information are filtered by type, and the target objects whose type belongs to the privacy object type set are determined as candidate privacy objects of the image to be processed.

3. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the feature matching module performs feature extraction on the candidate privacy object, generates an object feature vector of the candidate privacy object, and matches the object feature vector with a preset privacy feature library to obtain the matching confidence score of the candidate privacy object, it is specifically used for: Perform key point detection on the candidate privacy object and extract the key point location information of the candidate privacy object; Based on the key point location information, a local texture descriptor of the candidate privacy object is generated, and the local texture descriptor is combined into an object feature vector of the candidate privacy object; The object feature vector is compared with the reference feature vector stored in the preset privacy feature library to obtain the similarity between the object feature vector and the reference feature vector; The similarity is quantitatively evaluated with the similarity threshold of the candidate privacy object to obtain the matching confidence of the candidate privacy object.

4. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the mask generation module generates a privacy region mask for the candidate privacy object based on the object feature vector, the specific function of this module is as follows: Obtain the privacy attribute judgment result carried in the feature vector of the object; When the privacy attribute judgment result indicates that the candidate privacy object belongs to the privacy category, the spatial distribution information of the candidate privacy object is parsed from the object feature vector; Based on the spatial distribution information, locate the pixel region occupied by the candidate privacy object in the image coordinate system of the image to be processed; The pixel region is masked to obtain the privacy region mask of the candidate privacy object.

5. The image privacy content recognition and protection system as described in claim 4, characterized in that, When the mask generation module performs the judgment result of obtaining the privacy attribute carried in the feature vector of the object, it is specifically used for: Data structure parsing is performed on the object feature vector to extract feature fields carrying privacy category information from the object feature vector; The feature field is compared with the privacy category feature template recorded in the privacy category mapping table of the candidate privacy object to determine the privacy category matched by the feature field; When the feature field matches any privacy category feature template in the privacy category mapping table, the privacy attribute judgment result is determined as privacy; When the feature field does not match any of the privacy category feature templates in the privacy category mapping table, the privacy attribute judgment result is determined to be non-privacy.

6. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the level classification module performs the process of dividing the privacy region corresponding to the privacy region mask into different sensitivity levels based on the matching confidence level, and matching a corresponding privacy processing strategy for each sensitivity level, it is specifically used for: The matching confidence level is compared with a preset confidence threshold to determine the confidence interval to which the matching confidence level belongs; Based on the mapping relationship between the confidence interval and the sensitivity level, the sensitivity level of the privacy region corresponding to the privacy region mask is determined; Obtain the correspondence between the sensitivity level and the privacy processing strategy, and select a privacy processing strategy that matches the sensitivity level from the privacy processing strategy set according to the correspondence; The selected privacy processing strategy is determined as the privacy processing strategy corresponding to the privacy region mask.

7. The image privacy content recognition and protection system as described in claim 6, characterized in that, When the level classification module selects a privacy processing strategy matching the sensitivity level from the privacy processing strategy set according to the correspondence, it is specifically used for: Collect the mapping relationship between the sensitivity level and the privacy processing strategy; Based on the mapping relationship, the sensitivity level is semantically mapped and transformed to obtain the policy feature descriptor corresponding to the sensitivity level; Using the policy feature descriptor as the retrieval basis, feature matching is performed on the privacy processing policy set to obtain the privacy processing policy that matches the policy feature descriptor.

8. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the privacy processing module performs differential privacy processing on the image region covered by the privacy region mask according to the privacy processing strategy to generate an intermediate image of the candidate privacy object, it is specifically used for: Extract the set of pixel coordinates corresponding to the mask identifier value in the privacy region mask. The set of pixel coordinates is used to indicate the pixel position occupied by the privacy region in the image to be processed. The privacy processing operation type of the candidate privacy object is parsed from the privacy processing strategy; Based on the privacy processing operation type, the pixel positions indicated by the pixel coordinate set are reconstructed and filled to obtain the processed privacy region image block of the candidate privacy object; The pixel positions indicated by the set of pixel coordinates are replaced with the processed privacy region image blocks to generate an intermediate image of the candidate privacy object.

9. The image privacy content recognition and protection system as described in claim 8, characterized in that, When the privacy processing module performs content reconstruction and filling on the pixel positions indicated by the pixel coordinate set according to the privacy processing operation type to obtain the processed privacy region image block of the candidate privacy object, it is specifically used for: When the privacy processing operation type is indicated as content reconstruction and filling, the image content reconstruction process for the pixel positions indicated by the pixel coordinate set is triggered; The spatial topology of the pixel coordinate set in the image to be processed is analyzed, and the neighborhood association relationship of the pixel coordinate set is determined based on the spatial topology. Based on the neighborhood association relationship, texture primitive features that are not affected by privacy processing are extracted around the set of pixel coordinates in the image to be processed. The texture primitive features are used to characterize the visual structure pattern of the neighborhood image region. Using the texture primitive features as the basis for reconstruction, a feature transfer operation is performed on the pixel positions indicated by the pixel coordinate set to generate reconstructed pixel content that matches the visual features in the neighboring image region. The reconstructed pixel content is filled into the pixel positions indicated by the pixel coordinate set to obtain the processed privacy region image block of the candidate privacy object.

10. The image privacy content recognition and protection system as described in claim 1, characterized in that, When the image fusion module performs image fusion of the unprocessed region and the processed privacy region in the intermediate image to generate the output image of the candidate privacy object, it is specifically used for: Based on the coverage area of ​​the privacy region mask, the processed privacy region and the unprocessed region are separated from the intermediate image; At the boundary between the processed privacy region and the unprocessed region, extract the boundary pixel sequence of the processed privacy region and the adjacent pixel sequence of the unprocessed region; Using the boundary pixel sequence and the adjacent pixel sequence as references, pixel transition processing is performed on the processed privacy region and the unprocessed region to obtain the transition fusion band of the candidate privacy object; The transition fusion band is stitched together with the processed privacy region and the unprocessed region to obtain the output image of the candidate privacy object.

Images