Sensitive target detection method and device, equipment, medium and program product

Abstract

The invention provides a sensitive target detection method which can be applied to the technical field of artificial intelligence. The sensitive target detection method comprises the following steps: acquiring a target image; processing the target image through a target detection model to obtain a sensitive target detection result; wherein the target detection model comprises a dynamic depth separable convolution module, and the dynamic depth separable convolution module comprises a plurality of deep convolutions with different convolution kernel shapes; and respectively extracting local features of the input features of the dynamic depth separable convolution module by using the deep convolution of a plurality of different convolution kernel shapes, and carrying out weighted fusion on the local features to obtain features after feature extraction. The invention further provides a sensitive target detection device and equipment, a storage medium and a program product.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, specifically to the field of image detection, and more specifically to a method, apparatus, device, medium, and program product for detecting sensitive targets. Background Technology

[0002] With the continuous development of the financial industry, more and more financial institutions are shifting their business online. In online identity verification and credit risk assessment, users typically need to upload materials such as headshots, front-facing photos, and / or full-body photos. The system extracts user features based on this image information and uses these features to identify the user and / or assess their credit risk. Image target detection technology, capable of identifying and classifying key targets in images, is crucial in this process. By applying target detection technology, the system can identify "sensitive targets" that significantly impact a user's credit score, thereby achieving automated user feature profiling and credit risk assessment. However, since sensitive targets are usually small in images, and existing models lack targeted optimization designs for small target detection, the accuracy of target detection is affected. Therefore, how to effectively detect and identify sensitive small targets has become a major challenge in applying target detection technology. Summary of the Invention

[0003] In view of the above problems, this application provides sensitive target detection methods, apparatus, equipment, media and program products to improve the accuracy of small target detection.

[0004] According to a first aspect of this application, a sensitive target detection method is provided, comprising: acquiring a target image; processing the target image through a target detection model to obtain a sensitive target detection result; wherein the target detection model includes a dynamically separable convolutional module, the dynamically separable convolutional module including multiple depth convolutions with different kernel shapes; extracting local features of the input features of the dynamically separable convolutional module using the multiple depth convolutions with different kernel shapes respectively, and weighted fusing the local features to obtain the extracted features.

[0005] According to an embodiment of this application, the target detection model includes a dynamic feature extraction module; a first feature map of the input features of the dynamic feature extraction module is extracted using the convolutional layer of the dynamic feature extraction module, and the first feature map is split into a second feature map and a third feature map; the second feature map is extracted using n dynamic hybrid convolutional modules of the dynamic feature extraction module to obtain a fourth feature map; wherein n is an integer greater than 1, and the dynamic hybrid convolutional module includes a dynamic depthwise separable convolutional module; the first feature map, the third feature map and the fourth feature map are concatenated and fused to obtain a fifth feature map, and the fifth feature map is integrated using a convolutional fusion layer to obtain the output features of the dynamic feature extraction module.

[0006] According to an embodiment of this application, the dynamic hybrid convolution module includes a dynamic mixer, which includes a dynamic depthwise separable convolution module; the input features of the dynamic hybrid convolution module are segmented into channels, and the dynamic mixer uses dynamic depthwise separable convolution modules with different receptive fields to extract features from the segmented features, and the extracted features are concatenated and output.

[0007] According to an embodiment of this application, the dynamic depthwise separable convolutional module includes a weighted fusion branch and multiple feature extraction branches. Each feature extraction branch includes a depthwise convolutional layer, and the kernel shape of the depthwise convolutional layer of each feature extraction branch is different. The weighted fusion branch is used to perform global average pooling and convolution operations on the input features of the dynamic depthwise separable convolutional module to obtain the weights corresponding to each feature extraction branch, so as to weightedly fuse the local features extracted by each feature extraction branch.

[0008] According to an embodiment of this application, the target detection model further includes a local-global fusion module, which is located at the tail of the neck network of the target detection model. The local-global fusion module extracts the global attention features and local attention features of the input features of the local-global fusion module respectively, and concatenates and fuses the global attention features and local attention features to obtain the output features of the local-global fusion module.

[0009] According to an embodiment of this application, the local-global fusion module includes a local-global attention module; the local-global attention module is used to extract global attention features and local attention features; wherein, when extracting global attention features, the convolution kernel of the local-global attention module has a first size, and when extracting local attention features, the convolution kernel of the local-global attention module has a second size, and the first size is larger than the second size.

[0010] According to an embodiment of this application, the target detection model further includes a feature enhancement module, which is located at the tail of the backbone network of the target detection model. The feature enhancement module includes a sparse self-attention module. An attention enhancement feature is generated from the input feature of the sparse self-attention module using an adaptive sparse self-attention sub-module. The attention enhancement feature is then connected to the input feature residual of the sparse self-attention module to obtain the attention output feature. The attention output feature is then nonlinearly mapped using a feedforward neural network, and the result of the nonlinear mapping is added to the attention output feature through the residual connection to obtain the output feature of the sparse self-attention module.

[0011] A second aspect of this application provides a sensitive target detection apparatus, comprising: an image acquisition module for acquiring a target image; and a sensitive target detection module for processing the target image using a target detection model to obtain a sensitive target detection result; wherein the target detection model includes a dynamic depthwise separable convolution module, the dynamic depthwise separable convolution module including multiple depthwise convolutions with different kernel shapes; local features of the input features of the dynamic depthwise separable convolution module are extracted using the multiple depthwise convolutions with different kernel shapes, and the local features are weighted and fused to obtain the extracted features.

[0012] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0013] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0014] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0015] The above one or more embodiments have the following beneficial effects: In the embodiments of this application, the target detection model extracts features through a dynamic depth separable convolution module. The dynamic depth separable convolution module uses convolution kernels of different shapes in parallel, which can capture local information of multiple directions and scales in the image, thereby enhancing the model's ability to extract local information in the image. This can at least partially solve the problem of low detection accuracy of small targets in the traditional target detection process, and thus achieve the effect of improving the accuracy of sensitive target detection. Attached Figure Description

[0016] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0017] Figure 1 The illustration shows an application scenario of the sensitive target detection method, apparatus, device, medium, and program product according to embodiments of this application;

[0018] Figure 2 A flowchart illustrating a sensitive target detection method according to an embodiment of this application is shown schematically.

[0019] Figure 3 The schematic diagram illustrates the structure of a target detection model according to an embodiment of this application;

[0020] Figure 4 The schematic diagram illustrates the structure of a dynamic feature extraction module according to an embodiment of this application;

[0021] Figure 5 The schematic diagram illustrates the structure of a dynamic hybrid convolution module according to an embodiment of this application;

[0022] Figure 6 A schematic diagram of the structure of a dynamic mixer according to an embodiment of this application is shown.

[0023] Figure 7 The schematic diagram illustrates the structure of a dynamically depth-separable convolutional module according to an embodiment of this application;

[0024] Figure 8 The schematic diagram illustrates the structure of a local-to-global fusion module according to an embodiment of this application;

[0025] Figure 9 This illustration schematically shows a structural diagram of a local global attention module according to an embodiment of this application;

[0026] Figure 10 The schematic diagram illustrates the structure of a feature enhancement module according to an embodiment of this application;

[0027] Figure 11 The schematic diagram illustrates the structure of a sparse self-attention module according to an embodiment of this application;

[0028] Figure 12 The schematic diagram illustrates the structure of a sparse self-attention submodule according to an embodiment of this application;

[0029] Figure 13 This schematically illustrates a structural block diagram of a sensitive target detection device according to an embodiment of this application; and

[0030] Figure 14 A block diagram schematically illustrates an electronic device suitable for implementing a sensitive target detection method according to an embodiment of this application. Detailed Implementation

[0031] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0032] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0033] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0034] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0035] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0036] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0037] Figure 1 The illustration shows an application scenario of the sensitive target detection method according to an embodiment of this application.

[0038] like Figure 1As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0039] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0040] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers. Furthermore, the first terminal device 101, the second terminal device 102, and the third terminal device 103 can also be business processing terminals with image acquisition capabilities at the business site.

[0041] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0042] It should be noted that the sensitive target detection method provided in this application embodiment can generally be executed by server 105. Correspondingly, the sensitive target detection device provided in this application embodiment can generally be located in server 105. The sensitive target detection method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the sensitive target detection device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0043] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0044] The following will be based on Figure 1 The described scene, through Figures 2-12 A sensitive target detection method according to embodiments of this application will be described in detail.

[0045] Figure 2 A flowchart illustrating a sensitive target detection method according to an embodiment of this application is shown.

[0046] like Figure 2 As shown, the sensitive target detection method in this embodiment includes operations S210 to S220.

[0047] In operation S210, the target image is acquired.

[0048] In embodiments of this application, the target image may include a user image containing user information. When applied to sensitive target detection in other scenarios, the target image may further include non-user images, such as environmental images.

[0049] Before obtaining user information, obtain the user's consent or authorization. For example, before operation S210, if the target image includes a user image, a request to obtain the target image can be sent to the user, and the user's authorization for obtaining the target image can be obtained. If the user agrees or authorizes the acquisition of the target image, operation S210 is executed to acquire the target image.

[0050] In the embodiments of this application, the target images obtained include, but are not limited to, user ID photos, user on-site photos, or historical archived images. By accurately identifying the target images, the user's identity can be identified, and the user's risk coefficient can be further determined based on the identification results, thereby effectively preventing and controlling credit risk.

[0051] In operation S220, the target image is processed by the target detection model to obtain the sensitive target detection result.

[0052] In the embodiments of this application, a corresponding operation entry can be provided to the user, allowing the user to choose to agree to or reject the automated decision result. That is, before performing sensitive target detection on the target image, the user can provide an instruction to agree to or reject sensitive target detection through the corresponding operation entry. If the user agrees to perform sensitive target detection, then sensitive target detection is performed on the target image, i.e., step S220 is executed. If the user refuses to perform detection, then the expert decision-making process is initiated.

[0053] In the embodiments of this application, the object detection model is a pre-trained neural network model. Addressing the shortcomings of traditional neural network models in extracting local information for small objects, the object detection model uses depthwise convolutions with different kernel shapes in parallel, such as square kernels, horizontal bar kernels, and vertical bar kernels, to capture local information from multiple directions and scales in the image. The extracted local information is further fused using a weighted fusion method, thereby achieving sufficient feature extraction and accurate detection, reducing false positives and false negatives for sensitive small objects, and mitigating the problem of missed detections due to the small size of the target.

[0054] Figure 3 The schematic diagram illustrates the structure of a target detection model according to an embodiment of this application, such as... Figure 3 As shown in the embodiments of this application, the target detection model includes a backbone network, a neck network, and a head network, specifically including a convolution module, a dynamic feature extraction module, a fast spatial pyramid pooling module, a feature enhancement module, an upsampling module, a stitching module, a local-global fusion module, and a detection head module.

[0055] It should be noted that, in order to clearly describe the connection relationship between the modules, the modules in the embodiments of this application are distinguished by "first", "second", etc. For example, "first dynamic feature extraction module" and "second dynamic feature extraction module" are dynamic feature extraction modules located in different positions, and their structures can be the same or different. In the embodiments of this application, the structures of each dynamic feature extraction module are the same.

[0056] In the embodiments of this application, the backbone network is used to implement multi-scale feature extraction, including a convolution module, a first dynamic feature extraction module, a second dynamic feature extraction module, a third dynamic feature extraction module, a fourth dynamic feature extraction module, a fast spatial pyramid pooling module, and a feature enhancement module connected in sequence.

[0057] The neck network is used to achieve multi-scale feature fusion, including a first upsampling module, a first concatenation module, a fifth dynamic feature extraction module, a second upsampling module, a second concatenation module, a sixth dynamic feature extraction module, a convolution module, a third concatenation module, a seventh dynamic feature extraction module, a convolution module, a fourth concatenation module, and an eighth dynamic feature extraction module, as well as a first local-global fusion module, a second local-global fusion module, and a third local-global fusion module for achieving feature fusion. Specifically, the first upsampling module upsamples the output features of the feature enhancement module and inputs them into the first concatenation module; the first concatenation module concatenates the output features of the first upsampling module and the output features of the third dynamic feature extraction module and inputs them into the fifth dynamic feature extraction module; the second upsampling module upsamples the output features of the fifth dynamic feature extraction module and inputs them into the second concatenation module; the second concatenation module concatenates the output features of the second upsampling module and the output features of the second dynamic feature extraction module and inputs them into the sixth dynamic feature extraction module; the third concatenation module concatenates the output features of the fifth dynamic feature extraction module and the output features of the sixth dynamic feature extraction module after processing by the convolution module and inputs them into the seventh dynamic feature extraction module; and the fourth concatenation module concatenates the output features of the feature enhancement module and the output features of the seventh dynamic feature extraction module after processing by the convolution module and inputs them into the eighth dynamic feature extraction module. The first local-to-global fusion module integrates the output features of the sixth dynamic feature extraction module and the eighth dynamic feature extraction module, and inputs them into the second local-to-global fusion module and the head network; the second local-to-global fusion module integrates the output features of the seventh dynamic feature extraction module and the first local-to-global fusion module, and inputs them into the head network and the third local-to-global fusion module; the third local-to-global fusion module integrates the output features of the eighth dynamic feature extraction module and the second local-to-global fusion module, and inputs them into the head network.

[0058] The head network is used to generate feature detection results and includes a first detection head module, a second detection head module, and a third detection head module. The first detection head module receives the output features from the first local-global fusion module and performs feature detection; the second detection head module receives the output features from the second local-global fusion module and performs feature detection; and the third detection head module receives the output features from the third local-global fusion module and performs feature detection. By fusing the feature detection results from the first, second, and third detection head modules, the output result of the target detection module can be obtained.

[0059] In the embodiments of this application, a dynamic depthwise separable convolution module is designed for the small target detection task to improve the accuracy of small target detection. Specifically, in order to improve the detection efficiency on sensitive targets, the dynamic feature extraction module in the embodiments of this application includes a dynamic depthwise separable convolution module. The dynamic depthwise separable convolution module extracts local features of the input features through depthwise convolution with different kernel shapes, and weights and fuses the local features to obtain the extracted features. This effectively captures local information of multiple directions and scales in the image, improving the ability to detect small targets and extract features.

[0060] The dynamic depthwise separable convolution module can directly adopt a feature extraction structure with multiple parallel convolutional kernels and add it to the model as a complete dynamic feature extraction module. For example, three parallel depthwise convolutions with different kernel shapes can be used as feature extraction branches, and an adaptive fusion module can be added after each branch to perform weighted fusion of the features extracted by different branches. Then, the output features are obtained after post-processing (such as normalization and activation functions). Alternatively, the dynamic depthwise separable convolution module can be used as a sub-module of the dynamic feature extraction module, and other data processing modules can be added before and after this sub-module to achieve better small object detection results.

[0061] Figure 4 A schematic diagram illustrating the structure of a dynamic feature extraction module according to an embodiment of this application is shown. Figure 4 As shown, the core processing structure of the dynamic feature extraction module consists of n stacked dynamic hybrid convolutional modules. The first feature map of the input features of the dynamic feature extraction module is extracted using the convolutional layers of the module. This first feature map is then split into a second and a third feature map using a splitting operation. The second feature map is then processed by the n dynamic hybrid convolutional modules to obtain a fourth feature map, where n is an integer greater than 1. The first, third, and fourth feature maps are then concatenated and fused to obtain a fifth feature map. Finally, a convolutional fusion layer integrates the features to obtain the output features of the dynamic feature extraction module. The dynamic feature extraction module can extract features at different scales, effectively enhancing the detection capability for small targets. It can also flexibly adjust its internal processing logic according to the distribution characteristics of sensitive targets in financial images, mitigating the problem of missed detections due to excessively small targets. This improves the robustness and detection accuracy of the model in image risk assessment scenarios, meeting the actual needs of financial institutions for accurate detection of sensitive targets in user images for credit risk assessment.

[0062] In the embodiments of this application, the dynamic hybrid convolution module combines adaptive depthwise convolution with efficient feature fusion. In the scenario of sensitive target detection in financial images, sensitive targets in financial images are usually small, making it difficult for traditional detection modules to effectively identify them. The adaptive depthwise convolution and efficient feature fusion of the dynamic hybrid convolution module enable it to accurately extract features under complex image backgrounds (such as uneven lighting, image blur, and other interference), reducing false detections and missed detections of sensitive small targets.

[0063] In the embodiments of this application, the dynamic hybrid convolution module employs a normalization layer, a dynamic mixer, and gated convolutional units to extract features from the input features. The dynamic mixer extracts local features through depthwise convolutions with different kernel shapes. The normalization layer standardizes the input tensor for easier subsequent data processing; the dynamic mixer extracts local features through depthwise convolutions with different kernel shapes, improving the model's feature extraction capability; and the gated convolutional units control the information flow.

[0064] In some implementations, gated convolutional units can be stacked in multiple layers to further enhance the expressive power of the model.

[0065] Figure 5 The schematic diagram illustrates the structure of a dynamic hybrid convolution module according to an embodiment of this application, such as... Figure 5 As shown in the embodiments of this application, the dynamic hybrid convolution module is a block structure that combines normalization, dynamic mixer, and gated convolution mechanisms. Its purpose is to enhance the model's ability to interact with information between local and global features. The dynamic hybrid convolution module first normalizes the input features through a normalization layer, then processes them sequentially through a dynamic mixer and gated convolution units. Next, it uses a multilayer perceptron (MLP) for feature mixing and modeling, and enhances the model's generalization ability through a path dropout mechanism. Finally, it obtains the output features through residual connections. Furthermore, this embodiment introduces a layer scaling weight parameter after the path dropout processing to adjust the strength of the residual connections.

[0066] In the embodiments of this application, the dynamic mixer can directly adopt a feature extraction structure with multiple parallel convolutional kernels to extract different local features, or it can be further designed as a feature extraction structure that can realize multi-scale feature extraction to further improve the model's ability to perceive local details and global context. Figure 6 A schematic diagram of the structure of a dynamic mixer according to an embodiment of this application is shown, such as... Figure 6As shown in the embodiments of this application, the dynamic mixer performs channel segmentation on the input features of the dynamic mixing convolution module and then uses dynamic depth separable convolution modules with different receptive fields to extract features. The extracted features are then concatenated and output. The dynamic depth separable convolution module includes multiple feature extraction branches, each feature extraction branch includes a depth convolution layer, and the convolution kernel shape of the depth convolution layer of each feature extraction branch is different.

[0067] In the embodiments of this application, the dynamic mixer is a dynamic multi-scale feature extraction unit that combines dynamically deep separable convolutional modules of different sizes to enhance the model's ability to perceive local details and global context. This structure groups information along the channel dimension and extracts features in parallel through convolutions with different receptive fields, thereby capturing multi-scale information and ultimately... Convolutional integration of feature output. Compared to traditional parallel convolutional modules, this structure enhances the diversity and dynamic adaptability of feature representation while maintaining lightweight design. Its specific implementation process is as follows: First, the input features are evenly divided into two sub-feature maps along the channel dimension: a first sub-feature map and a second sub-feature map, each with dimension R. Then, dynamic depthwise separable convolutions with different kernel sizes are applied to the two sub-feature maps for multi-scale feature extraction. In the embodiment of this application, the first sub-feature map is input to a kernel with a size of R. In the dynamic depthwise separable convolution, the first output is obtained; the second sub-feature map is input into the kernel with a size of In the dynamically depthwise separable convolution, the second output is obtained; then the first and second outputs are concatenated along the channel dimension to merge them into a new feature map; finally, through a... Convolution performs channel blending and integration on the concatenated feature maps to obtain the final output features of the module.

[0068] Figure 7 The schematic diagram illustrates the structure of a dynamically depth-separable convolutional module according to an embodiment of this application. See also... Figure 7The dynamic depthwise separable convolution module includes multiple feature extraction branches, each of which includes a depthwise convolutional layer with different kernel shapes. In this embodiment, the dynamic depthwise separable convolution module includes a first feature extraction branch, a second feature extraction branch, and a third feature extraction branch. The first feature extraction branch has a depthwise convolution with a first kernel shape, the second feature extraction branch has a depthwise convolution with a second kernel shape, and the third feature extraction branch has a depthwise convolution with a third kernel shape. For example, the kernel shape of the depthwise convolution in the first feature extraction branch is square, the kernel shape of the depthwise convolution in the second feature extraction branch is horizontally striped (i.e., horizontal bar), and the kernel shape of the depthwise convolution in the third feature extraction branch is vertically striped (i.e., vertical bar). In other embodiments, for further local feature extraction, convolutional kernels of other shapes, such as rectangles, can also be used. The Dynamic Depth-Separable Convolution module implements dynamically weighted multi-branch depth-separable convolution, capturing local information in images at multiple directions and scales by using convolution kernels of different shapes (square, horizontal, vertical) in parallel.

[0069] See you again Figure 7 In the embodiments of this application, the dynamic depthwise separable convolution module further includes a weighted fusion branch. The weighted fusion branch performs global average pooling and convolution operations on the input features of the dynamic depthwise separable convolution module to obtain the weights corresponding to each feature extraction branch, so as to weightedly fuse the local features extracted by each feature extraction branch.

[0070] In the embodiments of this application, by using a dynamic attention mechanism to adaptively fuse the output of each convolutional branch, the fusion weights can be flexibly adjusted according to the distribution characteristics of sensitive targets in the financial image, thereby enhancing the model's adaptability to structural orientation.

[0071] In some implementations, other weighted fusion methods can also be used to achieve weighted fusion of local features obtained from multiple feature extraction branches. For example, based on a gating mechanism, multiple local features can be concatenated along the channel dimension and input into a small convolutional neural network or multilayer perceptron, outputting normalized weights corresponding to multiple branches, and then using the obtained weights to perform weighted fusion of each branch.

[0072] In the embodiments of this application, after weighted fusion is completed, the fusion result is further subjected to batch normalization and activation function processing to finally obtain the output features of the dynamic depthwise separable convolutional module. Taking a module that includes three feature extraction branches and one weighted fusion branch as an example, the processing flow of the module is as follows.

[0073] First, the input features of the dynamically depthwise separable convolutional module are fed into three depthwise separable convolutional paths with different kernel shapes, respectively: using... Convolution simulates a square local receptive field, and the output yields the first local feature. ;use Convolution simulates a lateral band-shaped receptive field, and the output yields the second local feature. ;use Convolution simulates a longitudinal strip-shaped receptive field, and the output yields the third local feature. Where B represents the batch size, C represents the number of channels, and H and W represent the height and width of the feature, respectively. Then, the three feature extraction branches are dynamically weighted, and global average pooling is used to extract global channel features from the input features. Finally, a... Convolution generates a shape of Dynamic weight tensor The values ​​of i=1,2,3 are then normalized to normalize the importance of each branch. Next, the output is dynamically fused: the output of each feature extraction branch is... Its corresponding weight Multiply the results, stack the weighted outputs along the dimensions, and sum them to obtain the fused result y. Finally, the fused result... After batch normalization and SiLU activation function processing, the output features are obtained. .

[0074] See you again Figure 3 In this embodiment, a local-global fusion module is further designed at the tail of the neck network of the object detection model to perform refined fusion of input features through hierarchical attention branch processing paths. The local-global fusion module employs a local-global attention module to extract semantic context information from both local and global scales and fuse information flows from the two input paths, thereby enhancing the ability to express structural, edge, and semantic features. It also effectively fuses features from different sources at the channel dimension, thereby improving the model's ability to understand structure, boundaries, and semantics in complex scenes.

[0075] Figure 8 The schematic diagram illustrates the structure of a local-to-global fusion module according to an embodiment of this application, such as... Figure 8 As shown, the local-global fusion module extracts semantic context information from both local and global scales and fuses information flows from two input paths, thereby enhancing the ability to express structural, edge, and semantic features. It also effectively fuses features from different sources at the channel dimension, thereby improving the model's ability to understand structure, boundaries, and semantics in complex scenes.

[0076] In the embodiments of this application, the local global attention module of the local global fusion module is used to extract global attention features and local attention features of the input features; when extracting global attention features, the convolution kernel of the local global attention module is of a first size, and when extracting local attention features, the convolution kernel of the local global attention module is of a second size, and the first size is larger than the second size.

[0077] The processing flow of the local-to-global fusion module includes: First, multiple (two in this embodiment) input feature maps from different levels or paths are processed. Through respectively Convolution performs channel compression, unifying the channel dimensions and reducing computational complexity to obtain compressed feature representations. Next, respectively... and Multi-scale attention modeling is performed: context enhancement is achieved using local and global attention modules. The local attention module employs a smaller receptive field, and the convolutional kernels of the local and global attention modules are set... It emphasizes spatial details and is suitable for detecting edges and textured regions of sensitive targets; the global attention module uses a large receptive field, and the convolution kernel of the local global attention module is set to... Extract the overall structural semantics. (This is done separately for...) and After extracting these two attention features, they are concatenated along the channel dimension to obtain the enhanced feature vectors. .Will and Adding them together represents the fused semantic information of the two, and is done through a... The convolution process further extracts features, yielding the extracted feature vector b, which supplements the final feature alignment. Subsequently, the three sub-branches... By splicing and combining along the channel dimension, a fusion feature is formed. This feature is first achieved through a Convolution performs channel compression, followed by a reparameterizable convolution with a group convolutional structure to enhance feature representation capabilities, and finally through a... Convolution is mapped to the output dimension to generate the final fused features. .

[0078] Figure 9 The schematic diagram illustrates the structure of a local / global attention module according to an embodiment of this application, such as... Figure 9As shown, the Local-Global Attention module enhances the modeling ability of feature representations for details and overall structure by introducing a two-level attention mechanism (local and global attention by channel) from the previous level. This module receives a two-dimensional spatial feature map as input and uses a weighted mechanism of local context and global semantics to enhance features, ultimately outputting a spatially consistent, channel-enhanced feature map. The process of the Local-Global Attention module includes: first, expanding the input features into local patch tensors according to patch size; after average pooling of each patch, inputting it into the first fully connected network MLP1 for dimensionality reduction encoding; then normalizing it and using the second fully connected network MLP2 to restore the channel dimension; subsequently, obtaining the local attention distribution through Softmax normalization and multiplying it with itself to obtain a weighted feature representation; finally, calculating its cosine similarity to the learnable semantic prompt, obtaining the value range within... An attention mask is applied to the original feature map to enhance semantically relevant regions and suppress irrelevant regions. Finally, the feature map is channel-mapped using a learnable linear transformation matrix, upsampled back to the original spatial dimension, and then input. The final feature result after convolution output enhancement.

[0079] See you again Figure 3 In this embodiment, considering that sensitive targets in financial images are not only small in size but also unevenly distributed, a feature enhancement module is designed. By introducing an adaptive sparse self-attention mechanism, a refined enhancement of high-dimensional semantics is achieved, thereby improving the modeling capability of key regions.

[0080] Figure 10 The schematic diagram illustrates the structure of a feature enhancement module according to an embodiment of this application, such as... Figure 10 As shown, the feature enhancement module based on the improved adaptive sparse self-attention mechanism achieves refined enhancement of high-dimensional semantics by introducing multiple layers of adaptive sparse self-attention modules on the channel branches. The core process of this module is as follows: First, the input feature map is mapped to double the number of hidden channels through a dimensionality-reducing convolution, and is equally divided into two branches in the channel dimension: a main branch that retains the original information, and an attention branch to be enhanced. Then, the features of the attention branch are processed sequentially by m concatenated sparse self-attention modules to extract key contextual features of sparse attention. Finally, the main branch and the attention branch are concatenated in the channel dimension and fused through a dimensionality-increasing convolution to obtain the output features, thereby achieving the unification of cross-channel information interaction and sparse attention focusing.

[0081] The sparse self-attention module uses an adaptive sparse self-attention submodule to generate attention-enhanced features from the input features, and adds them to the input features through residual connections to obtain the attention output features. The attention output features are then subjected to nonlinear mapping using a feedforward neural network, and the result of the nonlinear mapping is added to the attention output features through residual connections to obtain the output features.

[0082] In the embodiments of this application, by introducing an adaptive sparse self-attention mechanism, a refined enhancement of high-dimensional semantics is achieved, thereby improving the model's ability to model key regions.

[0083] Figure 11 The schematic diagram illustrates the structure of a sparse self-attention module according to an embodiment of this application, as shown below. Figure 11 As shown, the sparse self-attention module introduces an adaptive sparse self-attention submodule and a feedforward network, combined with residual connections, to achieve sparse attention focusing and nonlinear transformation of input features. The processing flow of the sparse self-attention module includes: first, applying the adaptive sparse self-attention submodule to the input feature map to generate attention-enhanced features, and then adding them to the original input through residual connections to retain the original information while highlighting sparsely related features; based on the attention output, further performing nonlinear mapping through two layers of linear transformation and activation of the feedforward network, and then applying residual connections again to the features after the previous residual connection to enhance expressive power before outputting the final output feature map.

[0084] In the embodiments of this application, the adaptive sparse self-attention submodule is designed based on the adaptive sparse self-attention mechanism, based on the windowed multi-head self-attention framework, and combined with optional sparse mask and moving window strategy to take into account both local details and cross-window global information. Figure 12 The schematic diagram illustrates the structure of a sparse self-attention submodule according to an embodiment of this application, as follows: Figure 12 As shown, the adaptive sparse self-attention submodule first flattens the input feature map into a sequence and performs a first layer normalization to prepare for attention computation; secondly, it determines whether to perform cyclic translation on the sequence to break the window boundaries according to the configuration, and then performs the translation according to the window size. The features are divided into several non-overlapping sub-blocks. Then, within each sub-block, the similarity of query, key, and value is calculated using a multi-head mechanism. A sparse mask is applied to the similarity matrix based on the sparse attention flag, retaining only the positions with high relevance to the query, thereby reducing redundant computation. Next, the attention weights are Softmax normalized and multiplied by the value matrix to obtain the window output. Then, all window outputs are inversely merged and the cyclic translation is restored when the moving window is enabled. Then, the attention output is added to the original input residual through random path discarding, and then passed through a second layer of normalization and an optional feedforward network, with residual connections and random path discarding applied in the same way to complete the nonlinear feature transformation. Finally, the features in sequence form are reshaped into a spatial feature map, so that the size of the output feature map is consistent with that of the input feature map.

[0085] Compared with existing technologies, the target detection model proposed in this application for the application scenario of identifying sensitive targets not only enhances the model's ability to identify sensitive targets but also improves detection accuracy, meeting the actual needs of target identification in the financial field. Combined with computer algorithms, it can detect sensitive targets in target images in real time and simultaneously assess risk coefficients, reducing the intervention of manual screening and improving the efficiency of user risk assessment in the financial field.

[0086] In financial images, "sensitive small targets" such as ID card edges, watermarks, and tiny labels are often located at image edges or in the background, and are easily ignored by standard convolution. The designed feature extraction module, through adaptive depthwise convolution and efficient feature fusion, enables it to accurately extract features even in complex image backgrounds (such as uneven lighting and image blur), reducing false positives and false negatives for sensitive small targets. It also effectively enhances the detection capability for small targets and can flexibly adjust its internal processing logic according to the distribution characteristics of sensitive targets in financial images, mitigating the problem of false negatives caused by excessively small targets. This improves the model's robustness and detection accuracy in financial credit image risk assessment scenarios, meeting the actual needs of financial institutions for accurate detection of sensitive targets in target images for credit risk assessment.

[0087] In financial imagery, sensitive information often spans different scales; for example, facial features and full-body poses belong to different semantic levels. The designed local-global fusion module utilizes a hierarchical design, introducing local attention and global dependency modeling mechanisms to weightedly fuse feature maps from different depths and dynamically adjust feature importance through an attention mechanism. The local-global attention module extracts semantic contextual information from both local and global scales and fuses information flows from two input paths, thereby enhancing the ability to express structural, edge, and semantic features. It also effectively fuses features from different sources along the channel dimension, thus improving the model's understanding of structure, boundaries, and semantics in complex scenes.

[0088] Sensitive targets in financial images are not only small but also unevenly distributed. The designed feature enhancement module introduces an adaptive sparse self-attention mechanism to achieve refined enhancement of high-dimensional semantics, thereby improving the model's ability to model key regions.

[0089] In the embodiments of this application, the target detection model is a pre-trained neural network model. In practical applications, the model is first trained, and then the trained model is exported and deployed to perform sensitive target detection tasks.

[0090] First, a dataset was created by collecting images of sensitive targets from real-world financial business scenarios. This dataset, encompassing various image types including headshots, front-facing photos, and full-body shots, covers a range of sensitive targets such as gold chains, glasses, headphones, and tattoos.

[0091] Secondly, sensitive targets in the dataset are labeled, and data augmentation and enhancement are performed. Images are imported into a sensitive target labeling tool for labeling, and the labeling file contains the category of each sensitive target. Data augmentation methods such as random contrast enhancement, noise reduction, flipping, and scaling are used to augment the image data and labels, increasing the quantity to a set multiple of the original dataset. This simulates images recognized by the camera under various extreme conditions, thereby improving the generalization ability of the trained model.

[0092] The dataset is divided into validation, training, and test sets, and processed into a format that meets the model's input requirements. The dataset is then divided into training, validation, and test sets in a set ratio of 8:1:1. The training set will be used to train the model, the validation set will be used for evaluation during training, and the test set will be used to evaluate the model's performance.

[0093] The model is trained using training and validation sets. The training and validation sets are input into the object detection model, and the number of training iterations is set. As the number of training iterations increases, the loss function curve of the object detection model gradually converges. When the loss function curve converges and stabilizes, the model is trained to its optimal state. The optimal model weight file is saved. The images to be detected in the test set are input into the trained object detection model, and the sensitive object detection results are output. The detection results include the type of each detected object, and the location of each object in the image can be marked.

[0094] Finally, the optimal weight file is exported, and the trained model is used to monitor the target image in real time.

[0095] Based on the above-described sensitive target detection method, this application also provides a sensitive target detection device. The following will be combined with... Figure 13 The device is described in detail.

[0096] Figure 13 A schematic block diagram of a sensitive target detection device according to an embodiment of this application is shown.

[0097] like Figure 13 As shown, the sensitive target detection device 1300 of this embodiment includes an image acquisition module 1310 and a sensitive target detection module 1320.

[0098] The image acquisition module 1310 is used to acquire a target image. In one embodiment, the image acquisition module 1310 can be used to perform the operation S210 described above, which will not be repeated here.

[0099] The sensitive target detection module 1320 is used to process the target image using a target detection model to obtain the sensitive target detection result. In one embodiment, the sensitive target detection module 1320 can be used to perform the operation S220 described above, which will not be repeated here.

[0100] According to an embodiment of this application, the target detection model includes a dynamic depthwise separable convolution module, which includes multiple depthwise convolutions with different kernel shapes; local features of the input features of the dynamic depthwise separable convolution module are extracted using the multiple depthwise convolutions with different kernel shapes, and the local features are weighted and fused to obtain the extracted features.

[0101] According to embodiments of this application, any plurality of modules in the image acquisition module 1310 and the sensitive target detection module 1320 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least a portion of the functionality of one or more of these modules can be combined with at least a portion of the functionality of other modules and implemented in one module. According to embodiments of this application, at least one of the image acquisition module 1310 and the sensitive target detection module 1320 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the image acquisition module 1310 and the sensitive target detection module 1320 can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0102] Figure 14 A block diagram schematically illustrates an electronic device suitable for implementing a sensitive target detection method according to an embodiment of this application.

[0103] like Figure 14 As shown, an electronic device 1400 according to an embodiment of this application includes a processor 1401, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1402 or a program loaded from a storage portion 1408 into a random access memory (RAM) 1403. The processor 1401 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1401 may also include onboard memory for caching purposes. The processor 1401 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0104] RAM 1403 stores various programs and data required for the operation of electronic device 1400. Processor 1401, ROM 1402, and RAM 1403 are interconnected via bus 1404. Processor 1401 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 1402 and / or RAM 1403. It should be noted that programs may also be stored in one or more memories other than ROM 1402 and RAM 1403. Processor 1401 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0105] According to embodiments of this application, the electronic device 1400 may further include an input / output (I / O) interface 1405, which is also connected to a bus 1404. The electronic device 1400 may also include one or more of the following components connected to the input / output (I / O) interface 1405: an input section 1406 including a keyboard, mouse, etc.; an output section 1407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1408 including a hard disk, etc.; and a communication section 1409 including a network interface card such as a LAN card, modem, etc. The communication section 1409 performs communication processing via a network such as the Internet. A drive 1410 is also connected to the input / output (I / O) interface 1405 as needed. A removable medium 1411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1410 as needed so that computer programs read from it can be installed into the storage section 1408 as needed.

[0106] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0107] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 1402 and / or RAM 1403 and / or one or more memories other than ROM 1402 and RAM 1403 described above.

[0108] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the sensitive target detection method provided in the embodiments of this application.

[0109] When the computer program is executed by the processor 1401, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0110] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1409, and / or installed from the removable medium 1411. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0111] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1409, and / or installed from the removable medium 1411. When the computer program is executed by the processor 1401, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0112] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0113] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0114] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A method for detecting sensitive targets, characterized in that, include: Acquire the target image; The target image is processed by a target detection model to obtain the sensitive target detection result; The target detection model includes a dynamic depthwise separable convolution module, which includes multiple depthwise convolutions with different kernel shapes. Local features of the input features of the dynamic depthwise separable convolution module are extracted using the multiple depthwise convolutions with different kernel shapes, and the local features are weighted and fused to obtain the extracted features.

2. The method according to claim 1, characterized in that, The target detection model includes a dynamic feature extraction module; The first feature map of the input features of the dynamic feature extraction module is extracted using the convolutional layer of the dynamic feature extraction module, and the first feature map is split into a second feature map and a third feature map. The second feature map is used to extract features using n dynamically hybrid convolutional modules of the dynamic feature extraction module to obtain a fourth feature map; where n is an integer greater than 1, and the dynamically hybrid convolutional module includes the dynamically depth-separable convolutional module. The first feature map, the third feature map, and the fourth feature map are concatenated and fused to obtain the fifth feature map. The fifth feature map is then integrated using a convolutional fusion layer to obtain the output features of the dynamic feature extraction module.

3. The method according to claim 2, characterized in that, The dynamic hybrid convolution module includes a dynamic mixer, which in turn includes a dynamically depth-separable convolution module. The input features of the dynamic hybrid convolution module are segmented into channels. The dynamic mixer is used to extract features from the channel-segmented features using the dynamic depth separable convolution module with different receptive fields. The extracted features are then concatenated and output.

4. The method according to claim 3, characterized in that, The dynamic depthwise separable convolution module includes a weighted fusion branch and multiple feature extraction branches. Each feature extraction branch includes a depthwise convolutional layer, and the kernel shape of the depthwise convolutional layer in each feature extraction branch is different. The weighted fusion branch is used to perform global average pooling and convolution operations on the input features of the dynamic depthwise separable convolution module to obtain the weights corresponding to each feature extraction branch, so as to weightedly fuse the local features extracted by each feature extraction branch.

5. The method according to claim 1, characterized in that, The target detection model also includes a local-global fusion module, which is located at the tail of the neck network of the target detection model; The global attention features and local attention features of the input features of the local global fusion module are extracted using the local global fusion module, and then the global attention features and local attention features are concatenated and fused to obtain the output features of the local global fusion module.

6. The method according to claim 5, characterized in that, The local-to-global fusion module includes a local-to-global attention module; The global attention features and the local attention features are extracted using the local global attention module; wherein, when extracting the global attention features, the convolution kernel of the local global attention module has a first size, and when extracting the local attention features, the convolution kernel of the local global attention module has a second size, and the first size is larger than the second size.

7. The method according to claim 1, characterized in that, The target detection model also includes a feature enhancement module, which is located at the tail of the backbone network of the target detection model; The feature enhancement module includes a sparse self-attention module; Using the adaptive sparse self-attention submodule of the sparse self-attention module, attention enhancement features of the input features of the sparse self-attention module are generated; the attention enhancement features are concatenated with the residual of the input features of the sparse self-attention module to obtain attention output features; The attention output features are nonlinearly mapped using a feedforward neural network, and the result of the nonlinear mapping is added to the attention output features through residual connections to obtain the output features of the sparse self-attention module.

8. A sensitive target detection device, characterized in that, include: The image acquisition module is used to acquire the target image; as well as The sensitive target detection module is used to process the target image through a target detection model to obtain the sensitive target detection result; The target detection model includes a dynamic depthwise separable convolution module, which includes multiple depthwise convolutions with different kernel shapes. Local features of the input features of the dynamic depthwise separable convolution module are extracted using the multiple depthwise convolutions with different kernel shapes, and the local features are weighted and fused to obtain the extracted features.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images