Depth difference polarization coding network-based camouflage person detection method
Through the method based on the depth differential polarization encoding network, using image information in four polarization directions, multi-scale features are extracted and fused, and the problem of insufficient feature learning limitations and differences in the detection of camouflage personnel is solved, achieving a higher detection accuracy rate.
Patent Information
- Application Number
- CN202411848876.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-16
- Publication Date
- 2025-03-18
AI Technical Summary
When the existing camouflage personnel detection methods deal with the similar colors and textures of the camouflage personnel and backgrounds, it is difficult to effectively distinguish them, and rely on a single RGB image information, resulting in limitations in feature learning.
Using a method based on a depth differential polarization coding network, the image information of the four polarization directions is used to improve the difference between the camouflage personnel and the background through steps such as depth differential polarization feature coding, global feature extraction, multi-scale feature fusion and object detection.
The accuracy of detection of camouflage personnel is improved, the difference between camouflage personnel and background is enhanced, and the limitations of single RGB image information is solved.
Smart Images
Figure CN119649009A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of disguised person detection, and in particular to a disguised person detection method based on a deep difference polarization coding network. Background Art
[0002] Disguised person detection is a target detection method that can identify and locate people wearing camouflage patterns in images. Disguised person detection faces more challenges than traditional target detection because people wearing artificial camouflage patterns have strong similarity with background color, texture and other feature information. Disguised person detection has become a research hotspot in the field of camouflaged target detection in computer vision, and has important implications for applications such as medical image segmentation and military target detection.
[0003] Existing camouflaged target detection methods are mainly aimed at the detection of naturally camouflaged biological targets, and there is little research on camouflaged person detection methods. Existing camouflaged person detection methods rely more on color, texture and other feature information in RGB images to enhance the distinguishability of camouflaged persons from the background, but a single RGB image information source has limitations in learning the color and texture features that distinguish camouflaged persons. Therefore, inspired by the fact that polarization information can reflect the properties of different objects, the present invention uses polarization images to propose a camouflaged person detection method based on a deep difference polarization coding network. Summary of the invention
[0004] The present invention proposes a disguised person detection method based on a deep difference polarization coding network, which can at least solve one of the technical problems in the background technology.
[0005] To achieve the above object, the present invention adopts the following technical solutions: A method for detecting a disguised person based on a deep difference polarization coding network comprises the following steps: S1. Deep difference polarization feature encoding; S2. Global feature extraction; S3. Multi-scale feature fusion; S4. Disguised personnel target detection; In step S1, the original polarization images of the four polarization directions Input into the depth difference polarization feature encoding module, and output the depth difference polarization feature vector of the disguised person .
[0006] In step S2, input the depth difference polarization feature vector , after the global feature extraction module, we get global feature vectors of different scales .
[0007] In step S3, the global feature vector is input, and after passing through the multi-scale feature attention fusion module, the fused feature vector is obtained. .
[0008] In step S4, the fused feature vector is input , after passing through the detector module, the detection result of the disguised person in the polarization image is obtained: Category , frame center coordinates , border width With height .
[0009] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.
[0010] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.
[0011] It can be seen from the above technical scheme that the disguised person detection method based on the deep difference polarization coding network of the present invention, on the one hand, addresses the difficulties that the disguised person and the background color texture are similar, the edge information is blended, and the target scale is variable. On the other hand, based on the problem that the current disguised person detection method extracts and learns distinguishable features, there is a single information source limitation (only RGB image information). A method for accurately detecting disguised persons using deep difference polarization coding features is proposed.
[0012] Specifically, the beneficial effects of the present invention are: The present invention realizes the deep difference polarization feature coding learning of disguised persons. The proposed deep difference polarization feature module can effectively encode image information in non-orthogonal polarization directions through non-linear convolution operations, improve the utilization rate of polarization feature information of disguised persons in polarization images, enhance the difference between disguised persons and backgrounds, and solve the problem of limited single distinguishable features in learning disguised persons.
[0013] The present invention realizes the attention fusion of multi-scale global polarization features. The proposed multi-scale attention fusion module adopts the channel attention mechanism to fuse multi-scale global polarization features, and adds the channel attention sub-module between features of different scales as a horizontal connection of fusion features, thereby improving the redundancy brought by the traditional fusion method of multi-scale global polarization features, reducing the interference of background polarization features in nonlinear space, improving the accuracy of disguised person detection, and solving the difficulties of disguised persons with variable scales and fusion of edge information. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is an overall structural diagram of a disguised person detection method based on a deep difference polarization coding network of the present invention; Figure 2It is a composition structure diagram of the deep difference polarization coding network of the present invention; Figure 3 It is a composition structure diagram of the global feature extraction module of the present invention; Figure 4 It is a composition structure diagram of the multi-scale feature attention fusion module of the present invention; Figure 5 It is a structural diagram of the detector module of the present invention. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.
[0016] The present invention proposes a method for detecting disguised persons based on a deep difference polarization coding network. Figure 1 As shown in the figure, the method consists of a deep difference polarization feature encoding module, a global feature extraction module, a multi-scale feature attention fusion module and a detector module. Specifically, given 0 0 , 45 0 , 90 0 , 135 0 Original polarization images in four polarization directions , , , ; Four polarization direction images As the input of the deep difference polarization feature module, the deep difference polarization feature of the disguised person is obtained after passing through 5 layers of difference cascade polarization feature encoding layers. ; Then, Input into the global feature extraction module to extract the global features of disguised personnel at different scales ,Furthermore, the global features of three different scales are input into the multi-scale feature attention fusion module, and the features of different scales are fused through the channel attention mechanism to obtain the multi-scale fusion feature vector Finally, the fused feature vector is input into the detector module and the category parameters are output respectively. and position parameters, including the coordinates of the center point of the border , border width With height .
[0017] The depth difference polarization feature encoding module is mainly used to extract the depth difference polarization features between the disguised person and the background in the polarization image. Its structure is as follows Figure 2As shown in the figure, it consists of four layers of convolutional coding blocks and one layer of degree of freedom convolutional coding block. Among them, the first layer of convolutional coding block consists of 48 channels of 1×1 convolutional layer, batch normalization layer and ReLU activation function; the second layer of convolutional coding block consists of 96 channels of 1×1 convolutional layer, batch normalization layer and ReLU activation function; the third layer of convolutional coding block consists of 32 channels of 1×1 convolutional layer, batch normalization layer and ReLU activation function; the fourth layer of convolutional coding block consists of 16 channels of 1×1 convolutional layer, batch normalization layer and ReLU activation function; the degree of freedom convolutional coding block consists of variable channels of 1×1 convolutional layer, batch normalization layer and ReLU activation function. The connection method between the convolutional coding blocks adopts a partial layer dense cascade method, which is mainly reflected in: the output feature vector of the first layer convolutional coding block is superimposed with the output feature vector of the second layer convolutional coding block as the input of the third layer convolutional coding block; the output feature vector of the first layer convolutional coding block, the output feature vector of the second layer convolutional coding block and the output feature vector of the third layer convolutional coding block are superimposed as the input of the fourth layer convolutional coding block.
[0018] Among them, the 1×1 convolution layer uses a 1×1 convolution kernel with a channel number of The input feature vector is convolved, and the number of output channels is The eigenvector of:
[0019] In the formula, Indicates the number of channels The input feature vector is Indicates the number of channels The output feature vector of Refers to 1×1 convolution operation. Batch normalization layer refers to calculating the mean and variance on each sample batch, normalizing the activation weights of each layer, and ensuring that the activation weights of each layer output maintain a stable distribution. ReLU activation function is a rectified linear unit, with input , the output can be expressed as
[0020] In summary, the 1×1 convolutional coding block can be expressed as
[0021] refers to the 1×1 convolution operation, Representation layer normalization operation, Represents the ReLU activation function operation.
[0022] The global feature extraction module mainly learns the global features between the depth difference polarization features of the disguised personnel to enhance the robustness of target detection at different scales. Its structure is as follows Figure 3 As shown in the figure, it consists of three feature extractors of different scales. Among them, the small-scale feature extractor consists of three layers of 7×7 convolution blocks, each of which consists of a 7×7 convolution layer, a batch normalization layer, and a ReLU activation function; the number of convolution channels in the first layer of 7×7 convolution blocks is 512, the number of convolution channels in the second layer of 7×7 convolution blocks is 1024, and the number of convolution channels in the third layer of 7×7 convolution blocks is 256. The medium-scale feature extractor consists of three layers of 5×5 convolution blocks, each of which consists of a 5×5 convolution layer, a batch normalization layer, and a ReLU activation function; the number of convolution channels in the first layer of 5×5 convolution blocks is 512, the number of convolution channels in the second layer of 5×5 convolution blocks is 1024, and the number of convolution channels in the second layer of 5×5 convolution blocks is 256. The large-scale feature extractor consists of three layers of 3×3 convolutional blocks. The number of convolutional channels in the first layer of 3×3 convolutional blocks is 512, the number of convolutional channels in the second layer of 3×3 convolutional blocks is 1024, and the number of convolutional channels in the second layer of 3×3 convolutional blocks is 256.
[0023] The multi-scale feature attention fusion module mainly uses the channel attention mechanism to fuse global features of different scales and suppress the redundant interference between global features of different scales. Its structure is as follows Figure 4 As shown in the figure, it consists of two upsampling layers and two channel attention blocks. The upsampling layer is implemented by bilinear interpolation operation; the channel attention consists of a global average pooling layer, 2 fully connected layers, ReLU activation function, Sigmoid activation function and multiplication operation layer.
[0024] The detector module mainly outputs the target category of the disguised person in the polarization image and its position coordinates in the image. Its structure is as follows Figure 5 As shown in the figure, it consists of a category prediction branch and a position prediction branch. The category prediction branch consists of three layers of 3×3 convolution blocks and one layer of 1×1 convolution blocks; the position prediction branch consists of three layers of 3×3 convolution blocks and two layers of 1×1 convolution blocks of different branches.
[0025] Accordingly, the disguised person detection method based on the deep difference polarization coding network of the present invention includes the following steps: S1. deep difference polarization feature coding; S2. global feature extraction; S3. multi-scale feature fusion; S4. disguised person target detection.
[0026] In step S1, the original polarization images of the four polarization directions Input into the depth difference polarization feature encoding module, and output the depth difference polarization feature vector of the disguised person .
[0027] In step S2, input the depth difference polarization feature vector , after the global feature extraction module, we get global feature vectors of different scales .
[0028] In step S3, the global feature vector is input, and after passing through the multi-scale feature attention fusion module, the fused feature vector is obtained. .
[0029] In step S4, the fused feature vector is input , after passing through the detector module, the detection result of the disguised person in the polarization image is obtained: Category , frame center coordinates , border width With height .
[0030] When a polarization image detection device is used to collect a polarization image of a scene containing a disguised person , which contains 0 0 , 45 0 , 90 0 , 135 0 The four polarization direction images of Taking advantage of the fact that polarization information can characterize the differences between objects with different attributes, the scene polarization image is Input into the disguised person detection method based on deep difference polarization coding network to obtain the detection results of disguised persons in polarization images, including categories , detect the coordinates of the center point of the border , detect border width With height .
[0031] Practice has proved that compared with the case where RGB images are used as input, the detection accuracy of disguised persons is improved by 10.18% after using the disguised person detection method with images in four polarization directions as input.
[0032] Compared with the original polarization image as input, the detection accuracy of disguised persons is improved by about 7.52% after using four polarization direction images as input and encoding deep difference polarization features.
[0033] Compared with the traditional polarization features as input, the accuracy of disguised personnel detection is improved by about 3.89% when the features output by the deep difference polarization encoding module after the four polarization direction images are input are used for detection.
[0034] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.
[0035] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.
[0036] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any of the disguised person detection methods based on a deep difference polarization coding network in the above embodiments.
[0037] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above methods.
[0038] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.
[0039] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.
[0040] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
[0041] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for detecting disguised persons based on a deep difference polarization coding network, characterized in that: The following steps are involved: S1. Deep difference polarization feature encoding; S2. Global feature extraction; S3. Multi-scale feature fusion; S4. Detection of disguised personnel targets; In step S1, the original polarization images of the four polarization directions Input into the depth difference polarization feature encoding module, and output the depth difference polarization feature vector of the disguised person ; , , , Given 0 respectively 0 , 45 0 , 90 0 , 135 0 Original polarization images in four polarization directions; In step S2, input the depth difference polarization feature vector , after the global feature extraction module, we get global feature vectors of different scales ; In step S3, the global feature vector is input, and after passing through the multi-scale feature attention fusion module, the fused feature vector is obtained. ; In step S4, the fused feature vector is input , after passing through the detector module, the detection result of the disguised person in the polarization image is obtained: Category , frame center coordinates , border width With height .
2. According to claim 1, a disguised person detection method based on a deep difference polarization coding network is characterized in that: The deep difference polarization feature encoding module is used to extract the deep difference polarization features between the camouflaged personnel and the background in the polarization image; it consists of four layers of convolutional encoding blocks and one layer of degree of freedom convolutional encoding block; among them, the first layer of convolutional encoding block consists of a 1×1 convolutional layer with 48 channels, a batch normalization layer, and a ReLU activation function; the second layer of convolutional encoding block consists of a 1×1 convolutional layer with 96 channels, a batch normalization layer, and a ReLU activation function; the third layer of convolutional encoding block consists of a 1×1 convolutional layer with 32 channels, a batch normalization layer, and a ReLU activation function; the fourth layer of convolutional encoding block consists of a 1×1 convolutional layer with 16 channels, a batch normalization layer, and a ReLU activation function; the degree of freedom convolutional encoding block consists of a 1×1 convolutional layer with a variable number of channels, a batch normalization layer, and a ReLU activation function; The connection method between the convolutional coding blocks adopts the partial layer dense cascade method. The output feature vector of the first layer convolutional coding block is superimposed with the output feature vector of the second layer convolutional coding block as the input of the third layer convolutional coding block; the output feature vector of the first layer convolutional coding block, the output feature vector of the second layer convolutional coding block and the output feature vector of the third layer convolutional coding block are superimposed as the input of the fourth layer convolutional coding block.
3. The method for detecting disguised persons based on a deep difference polarization coding network according to claim 2, characterized in that: in, The 1×1 convolution layer uses a 1×1 convolution kernel with a channel number of The input feature vector is convolved, and the number of output channels is The eigenvector of: In the formula, Indicates the number of channels The input feature vector is Indicates the number of channels The output feature vector of Refers to the 1×1 convolution operation; The batch normalization layer calculates the mean and variance on each sample batch and normalizes the activation weights of each layer to ensure that the activation weights output by each layer maintain a stable distribution; The ReLU activation function is a rectified linear unit, with the input being , the output is expressed as In summary, a 1×1 convolutional coding block is represented as refers to the 1×1 convolution operation, Representation layer normalization operation, Represents the ReLU activation function operation.
4. The method for detecting disguised persons based on a deep difference polarization coding network according to claim 1, characterized in that: The global feature extraction module learns the global features between the depth difference polarization features of the disguised person to enhance the robustness of target detection at different scales; The global feature extraction module structure consists of three feature extractors of different scales; among them, the small-scale feature extractor consists of three layers of 7×7 convolution blocks, each of which consists of a 7×7 convolution layer, a batch normalization layer, and a ReLU activation function; the number of convolution channels in the first layer of 7×7 convolution blocks is 512, the number of convolution channels in the second layer of 7×7 convolution blocks is 1024, and the number of convolution channels in the third layer of 7×7 convolution blocks is 256; The mesoscale feature extractor consists of three layers of 5×5 convolutional blocks, each of which consists of a 5×5 convolutional layer, a batch normalization layer, and a ReLU activation function; the number of convolutional channels in the first layer of 5×5 convolutional blocks is 512, the number of convolutional channels in the second layer of 5×5 convolutional blocks is 1024, and the number of convolutional channels in the second layer of 5×5 convolutional blocks is 256; The large-scale feature extractor consists of three layers of 3×3 convolutional blocks. The number of convolutional channels in the first layer of 3×3 convolutional blocks is 512, the number of convolutional channels in the second layer of 3×3 convolutional blocks is 1024, and the number of convolutional channels in the second layer of 3×3 convolutional blocks is 256.
5. The method for detecting disguised persons based on a deep difference polarization coding network according to claim 1, characterized in that: The multi-scale feature attention fusion module uses the channel attention mechanism to fuse global features of different scales and suppress the redundant interference between global features of different scales. Its structure consists of two upsampling layers and two channel attention blocks; the upsampling layer is implemented by bilinear interpolation operation; the channel attention consists of a global average pooling layer, 2 fully connected layers, ReLU activation function, Sigmoid activation function and multiplication operation layer.
6. The method for detecting disguised persons based on a deep difference polarization coding network according to claim 1, characterized in that: The detector module outputs the target category of the disguised person in the polarization image and its position coordinates in the image; its structure consists of a category prediction branch and a position prediction branch, where the category prediction branch consists of three layers of 3×3 convolution blocks and one layer of 1×1 convolution blocks; the position prediction branch consists of three layers of 3×3 convolution blocks and two layers of 1×1 convolution blocks with different branches.