A texture-guided method and system for detecting camouflaged targets

Abstract

This invention discloses a texture-guided camouflage target detection method and system, comprising acquiring a test image and inputting it into a trained camouflage target detection model for processing and outputting camouflage target detection results; specifically: extracting feature maps of each level of the test image through a feature extractor and outputting them to a texture supervision module; outputting the first-layer feature map to a target texture learning module and the highest-layer feature map to a semantic fusion module; the target texture learning module extracting target texture feature maps based on the first-layer feature map and outputting them to the texture supervision module; the texture supervision module performing deep feature extraction at each level based on the target texture feature map and the feature maps of each level to obtain deep feature maps of each level; and the semantic fusion module obtaining fusion features at each level based on the highest-layer feature map and the deep feature maps of each level, performing semantic fusion of the deep feature maps of each level with the fusion features of the next higher level, and outputting camouflage target detection results.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to a texture-guided method and system for detecting camouflaged targets. Background Technology

[0002] In nature, some animals blend into their environment by using colors, textures, or physical structures similar to their surroundings to avoid detection by predators. This camouflage strategy not only visually deceives predators but also makes it difficult for existing machine vision algorithms to distinguish camouflaged targets from images. Therefore, camouflage target detection (COD) has become a significant challenge in machine vision. Solving this problem will benefit many applications of computer vision, such as camouflage engineering detection in the military, polyp segmentation in medicine, and disease control in agriculture.

[0003] COD (Catastrophic Object Detection) aims to segment animals or man-made objects from images that have similar colors or textures to the background. Early works addressed COD by using handcrafted features, analyzing information such as color, motion, texture, and convexity in the image to enhance the salience of camouflaged targets and identify locations where they differ from the background. However, these methods struggled to accurately segment camouflaged targets in complex scenes. In recent years, inspired by the achievements of deep learning methods in image pattern recognition and salient object detection (SOD), some works have provided datasets for camouflaged target recognition and used deep learning methods to solve the COD problem. Compared to traditional methods, deep learning-based COD methods can capture low-level details and high-level semantic information in images and accurately segment camouflaged targets, achieving outstanding recognition results.

[0004] Although deep learning methods have achieved good results in detecting camouflaged targets, most current camouflage recognition algorithms are based on an encoder-decoder structure that uses the ground truth of the image to supervise the output of the network and each decoding layer. However, the features extracted by each layer of the model are of different orders. Using ground truth to supervise the learning of all parts of the network will not be conducive to the learning of each layer of the network. In addition, the high-order features extracted by the network are helpful in locating camouflaged targets in the image, while low-order information is helpful in recovering target details. Viewing different levels of information in isolation will not help improve the segmentation accuracy of the model.

[0005] Therefore, how to provide a texture-guided camouflage target detection method and system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a texture-guided method and system for detecting camouflaged targets to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A texture-guided camouflage target detection method includes:

[0009] S1. Acquire the image to be tested and input it into the trained camouflage target detection model;

[0010] S2. Process the image to be tested and output the camouflage target detection results;

[0011] The specific content of S2 includes:

[0012] S21. Extract feature maps of each level of the input image to be tested through the feature extractor and output them to the texture supervision module. Output the first-level feature map to the target texture learning module and output the highest-level feature map to the semantic fusion module.

[0013] S22. The target texture learning module extracts the target texture feature map based on the first-layer feature map and outputs it to the texture supervision module;

[0014] S23. The texture supervision module performs deep feature extraction at each level based on the target texture feature map output by the target texture learning module and the feature maps at each level output by the feature extractor, and obtains the deep feature maps at each level and outputs them to the semantic fusion module.

[0015] S24. The semantic fusion module obtains the fusion features of each level based on the highest-level feature map output by the feature extractor and the depth feature maps of each level output by the texture supervision module. It then performs semantic fusion between the depth feature maps of each level and the fusion features of the next higher level to output the camouflage target detection results.

[0016] Preferably, the specific content of S22 for extracting the target texture feature map is as follows: using a 1×1 convolution to reduce the number of channels in the first layer feature map to C1 / 4, and using multi-scale convolution. Features at different scales are extracted from the first-layer feature map X1, and integrated using two ConvBRU layers to obtain multi-scale features. The ConvBRU layers include 3×3 convolution, BatchNormalization, ReLU, and bilinear interpolation upsampling. The feature map size is restored to the same size as the input test image I, resulting in the output target texture feature map T. out .

[0017] Preferably, the calculation process of the S22 target texture learning module is as follows:

[0018]

[0019]

[0020] in, For the input image I∈R 3×H×W The feature maps obtained after the feature extractor at each level This is the first layer feature map. To perform convolution operations on the input using a convolutional layer with a kernel width of 2i-1 and a channel count of C1 / 4, Concat(·) is a channel-scale concatenation operation.

[0021] Preferably, the specific content of deep feature extraction at each level in S23 is as follows: Channel dimensionality reduction and multi-scale convolution are performed on the feature maps output by the feature extractor at each level; ConvBRD operation is performed on the target texture feature map, which consists of 3×3 convolution, Batch Normalization, ReLU, and downsampling; then element-wise multiplication and convolution with the feature maps at each level X are performed. i Element-wise addition is performed to obtain the deep feature maps S at each level output by the semantic fusion module. i .

[0022] Preferably, the calculation process of the S23 texture supervision module is as follows:

[0023]

[0024] in, For the features of each level of the input image, T out The target texture feature map is output by the target texture learning module. ConvBRD is a 3×3 convolution, BatchNormalization, ReLU and downsampling.

[0025] Preferably, the specific content of the S24 semantic fusion module is as follows:

[0026] S241. Receive the highest-level feature map output by the feature extractor and the depth feature maps of each level output by the texture supervision module;

[0027] S242. After performing channel dimensionality reduction and downsampling operations on the highest-level feature map output by the feature extractor and the highest-level depth feature map output by the texture supervision module, channel-level merging is performed, and the highest-level fused features and segmentation map are output through four convolutional blocks.

[0028] S243. After performing channel dimensionality reduction and downsampling operations on the depth feature maps of each level output by the texture supervision module and the fusion features of the next higher level, channel-level merging is performed to achieve the fusion of information of different orders. Then, the fusion features and segmentation maps of each level are output through four convolutional blocks until the segmentation map of the first layer is obtained as the camouflage target detection result.

[0029] Preferably, the training content for the camouflage target detection model includes:

[0030] A dataset of images containing camouflaged targets is pre-collected and labeled to obtain the image I and the camouflaged target label GT for each sample;

[0031] The pre-collected image dataset containing camouflaged targets is divided into a training set and a test set;

[0032] The camouflage target detection model is trained using the training set;

[0033] The trained camouflage target detection model was tested using a test set.

[0034] Preferably, the optimization objective for training the camouflage target detection model is:

[0035] L = L GT +L Tex

[0036] Where GT is the camouflaged target label of the input image I, and L GT For the semantic fusion module outputting the BCE and IOU losses with the ground truth, L Tex The L1 loss between the output of the target texture learning module and the target texture map;

[0037]

[0038]

[0039]

[0040] Among them, GT (w,h) and These represent the camouflage target label and the pixels in the camouflage target detection result output by the model, respectively. W and H are the image width and height, and GT is the ground truth value. (w,h) ∈{0,1} represents whether the pixel belongs to the camouflaged target;

[0041]

[0042] in, To obtain the target texture feature map T output by the target texture learning module out The integrated feature map, T, is obtained using a single-channel 1×1 convolution. (w,h) For the target texture map.

[0043] Preferably, the method for obtaining the target texture map is as follows: Extract the texture feature map of the input image using Local Binary Pattern (LBP), and extract the target texture features based on the camouflaged target label (GT) to obtain the target texture map. Specifically:

[0044] T = LBP(I) * GT.

[0045] A texture-guided camouflage target detection system includes an image acquisition module and a trained camouflage target detection model;

[0046] The image acquisition module is used to acquire the image to be tested and input it into the trained camouflage target detection model;

[0047] The trained camouflage target detection model is used to process the image under test and output the camouflage target detection results;

[0048] The camouflage target detection model includes a feature extractor, a target texture learning module, a texture supervision module, and a semantic fusion module;

[0049] The feature extractor is used to extract feature maps of each level of the input image to be tested and output them to the texture supervision module. It outputs the first-level feature map to the target texture learning module and the highest-level feature map to the semantic fusion module.

[0050] The target texture learning module is used to extract the target texture feature map based on the first-layer feature map and output it to the texture supervision module.

[0051] The texture supervision module is used to perform deep feature extraction at each level based on the target texture feature map output by the target texture learning module and the feature maps at each level output by the feature extractor, and to obtain the deep feature maps at each level and output them to the semantic fusion module.

[0052] The semantic fusion module is used to obtain fused features at each level based on the highest-level feature map output by the feature extractor and the depth feature maps at each level output by the texture supervision module. It then performs semantic fusion between the depth feature maps at each level and the fused features at the next higher level to output the camouflaged target detection results.

[0053] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a texture-guided camouflage target detection method and system. It uses the texture information of the camouflage target to guide the training of the lower layers of the network, and fully integrates the target texture information and the high- and low-order features extracted by the network during network training. Specifically, the Target Texture Learning Module (TTM) supervises the lower layers of the network to learn the texture of the camouflage target in the image, thereby improving the quality of the features extracted by the lower layers of the network. The features learned by the TTM and the features extracted by the lower layers of the network are jointly input into subsequent layers of the network to promote network learning. Simultaneously, the texture features learned by the TTM, along with the outputs of each level of the feature extractor, serve as input to the TSM to guide the learning of the image restoration part. The SFM integrates the outputs of each level of the network in the image restoration part, fuses the semantic information of adjacent layers of the network, and obtains the camouflage detection result under the supervision of the GT, improving the model segmentation accuracy and achieving advanced camouflage target detection results. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0055] Figure 1 The attached figure is a schematic diagram of the camouflage target detection model provided by the present invention;

[0056] Figure 2 The attached figure is a schematic diagram of the target texture learning module provided by the present invention;

[0057] Figure 3 The attached figure is a schematic diagram of the texture supervision module provided by the present invention;

[0058] Figure 4 The attached figure is a schematic diagram of the semantic fusion module provided by the present invention;

[0059] Figure 5 The attached figure is a schematic diagram of the target texture image acquisition process provided by the present invention;

[0060] Figure 6 The attached figure is a schematic diagram of the experimental results of the present invention and the comparative method provided in the embodiments of the present invention;

[0061] Figure 7 The attached figure is a schematic diagram of the visual comparison results provided by an embodiment of the present invention;

[0062] Figure 8 The attached figure is a schematic diagram of the ablation experiment results provided in an embodiment of the present invention;

[0063] Among them, Feature Extractor is the feature extractor, TTM is the target texture learning module, TSM is the texture supervision module, and SFM is the semantic fusion module. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] This invention discloses a texture-guided method for detecting camouflaged targets, comprising:

[0066] S1. Acquire the image to be tested and input it into the trained camouflage target detection model;

[0067] S2. Process the image to be tested and output the camouflage target detection results;

[0068] The specific content of S2 includes, for example Figure 1 :

[0069] S21. Extract the feature maps of each level of the input image to be tested through the feature extractor and output them to the texture supervision module TSM. Output the first layer feature map to the target texture learning module TTM and the highest layer feature map to the semantic fusion module SFM.

[0070] S22. The Target Texture Learning Module (TTM) extracts the target texture feature map based on the first-layer feature map and outputs it to the Texture Supervision Module (TSM).

[0071] S23. The texture supervision module TSM performs deep feature extraction at each level based on the target texture feature map output by the target texture learning module TTM and the feature maps at each level output by the feature extractor, and obtains the deep feature maps at each level and outputs them to the semantic fusion module SFM.

[0072] S24. The Semantic Fusion Module (SFM) obtains the fusion features of each level based on the highest-level feature map output by the feature extractor and the depth feature maps of each level output by the Texture Supervision Module (TSM). It then performs semantic fusion between the depth feature maps of each level and the fusion features of the next higher level to output the camouflage target detection results.

[0073] To further implement the above technical solutions, such as Figure 2 The specific steps for extracting the target texture feature map in S22 are as follows: A 1×1 convolution is used to reduce the number of channels in the first-layer feature map to C1 / 4, and multi-scale convolution is used... Features at different scales are extracted from the first-layer feature map X1, and integrated using two ConvBRU layers to obtain multi-scale features. The ConvBRU layers include 3×3 convolution, BatchNormalization, ReLU, and bilinear interpolation upsampling. The feature map size is restored to the same size as the input test image I, resulting in the output target texture feature map T. out .

[0074] To further implement the above technical solution, the calculation process of the S22 target texture learning module TTM is as follows:

[0075]

[0076]

[0077] in, For the input image I∈R 3×H×W The feature maps obtained after the feature extractor at each level This is the first layer feature map. To perform convolution operations on the input using a convolutional layer with a kernel width of 2i-1 and a channel count of C1 / 4, Concat(·) is a channel-scale concatenation operation.

[0078] To further implement the above technical solutions, such as Figure 3 The specific content of deep feature extraction at each level of S23 is as follows: Channel dimensionality reduction and multi-scale convolution are performed on the feature maps output by the feature extractor at each level. ConvBRD operation is then performed on the target texture feature map. ConvBRD involves 3×3 convolution, Batch Normalization, ReLU, and downsampling, followed by element-wise multiplication and convolution with the feature maps at each level X. i Element-wise addition is performed to obtain the depth feature maps S for each level output by the TSM module. i .

[0079] To further implement the above technical solution, the calculation process of the S23 texture supervision module TSM is as follows:

[0080] TSM(X i ,T out )

[0081]

[0082] in, For the features of each level of the input image, T out The target texture feature map is output by the target texture learning module TTM. ConvBRD is a 3×3 convolution, BatchNormalization, ReLU and downsampling.

[0083] To further implement the above technical solutions, such as Figure 4 The specific content of the S24 semantic fusion module (SFM) is as follows:

[0084] S241. Receive the highest-level feature map output by the feature extractor and the depth feature maps of each level output by the texture supervision module TSM;

[0085] S242. After performing channel dimensionality reduction and downsampling operations on the highest-level feature map output by the feature extractor and the highest-level depth feature map output by the texture supervision module TSM, channel-level merging is performed, and after passing through four convolutional blocks, the highest-level fused features and segmentation map are output.

[0086] S243. After performing channel dimensionality reduction and downsampling operations on the depth feature maps of each level output by the texture supervision module TSM and the fusion features of the next higher level, channel-level merging is performed to achieve the fusion of information of different orders. Then, the fusion features and segmentation maps of each level are output by four convolutional blocks until the segmentation map of the first layer is obtained as the camouflage target detection result.

[0087] To further implement the above technical solution, the training content of the camouflage target detection model includes:

[0088] A dataset of images containing camouflaged targets is pre-collected and labeled to obtain the image I and the camouflaged target label GT for each sample;

[0089] The pre-collected image dataset containing camouflaged targets is divided into a training set and a test set;

[0090] The camouflage target detection model is trained using the training set;

[0091] The trained camouflage target detection model was tested using a test set.

[0092] To further implement the above technical solution, the optimization objective for training the camouflage target detection model is:

[0093] L = L GT +L Tex

[0094] Where GT is the camouflaged target label of the input image I, and L GT L represents the BCE and IOU loss between the output of the semantic fusion module SFM and the ground truth. Tex The L1 loss is the difference between the output of the target texture learning module TTM and the target texture map.

[0095]

[0096]

[0097]

[0098] Among them, GT (w,h) and These represent the camouflage target label and the pixels in the camouflage target detection result output by the model, respectively. W and H are the image width and height, and GT is the ground truth value. (w,h) ∈{0,1} represents whether the pixel belongs to the camouflaged target;

[0099]

[0100] in, To obtain the target texture feature map T from the output of the target texture learning module TTM outThe integrated feature map, T, is obtained using a single-channel 1×1 convolution. (w,h) For the target texture map.

[0101] To further implement the above technical solutions, such as Figure 5 The target texture map is obtained by: extracting the texture feature map of the input image using Local Binary Pattern (LBP), and extracting the target texture features based on the camouflaged target label (GT) to obtain the target texture map. Specifically:

[0102] T = LBP(I) * GT.

[0103] In this embodiment, three commonly used datasets in the COD field are used to verify the performance of the camouflaged target detection model: CAMO, CHAMELEON, and COD10k. The CAMO dataset contains a total of 1250 images, of which 1000 are divided into the training set (CAMO-Train) and 250 are used as the test set (CAMO-Test). CHAMELEON contains a total of 76 images. COD10k contains 5066 images, of which 3040 are used as the training set (COD10k-Train) and the remaining 2026 are used as the test set (COD10k-Test). The model is trained using both CAMO-Train and COD10k-Train, and the model performance is tested on CAMO-Test, CHAMELEON, and COD10k-Test, respectively.

[0104] To quantitatively evaluate model performance, this embodiment uses four commonly used COD methods as evaluation metrics: weight-F-measure. Structure-measure (S) α ),Enhanced-measure(E Φ The performance of each method in detecting camouflaged targets was evaluated using a combination of methods and MAE.

[0105] We compared our invention with nine state-of-the-art (SOTA) methods in the COD field, including SINet, CubeNet, TINet, TANet, PFNet, R-MGL, NCL&HIT, BASNet, and GRN. Figure 6 As can be seen, the performance of this invention on the CAMO and COD10k datasets is significantly better than other methods, with all four evaluation metrics being higher than other comparative methods. Figure 7 The visualization comparison results of the present invention and the other three methods in the comparative experiment show that, for most input images, compared with the other methods in the figure, the present invention can not only detect disguised targets that are not easy to be found in the image, but its segmentation results are also very close to the GT value, and the segmentation effect of the present invention is more delicate.

[0106] Furthermore, to analyze the effectiveness of the camouflaged target detection model of this invention in the COD task, this embodiment sets up an ablation experiment: (a) to verify the effectiveness of the target texture map supervising network learning in this paper, the ground truth of the input image is used to supervise the learning of the TTM module; (b) to verify the effectiveness of the TSM using the output of the TTM to guide the learning of each layer of the network, the part of the TSM that processes texture features is removed; (c) to verify the effectiveness of using the SFM to fuse the semantic features of each level of the network, each SFM only learns the output from the TSM of its own layer, resulting in the ablation experiment results as shown below. Figure 8 .

[0107] The ablation experiment results show that the target detection model of this invention uses TTM to learn the texture features of camouflaged targets and uses SFM to fuse semantic information at different levels, which effectively improves the COD performance of the model. In addition, TSM also improves the camouflaged target detection performance of the network to a certain extent.

[0108] A texture-guided camouflage target detection system includes an image acquisition module and a trained camouflage target detection model;

[0109] The image acquisition module is used to acquire the image to be tested and input it into the trained camouflage target detection model;

[0110] The trained camouflage target detection model is used to process the image under test and output the camouflage target detection results;

[0111] The camouflage target detection model includes a feature extractor, a target texture learning module (TTM), a texture supervision module (TSM), and a semantic fusion module (SFM).

[0112] The feature extractor is used to extract feature maps of each level of the input image to be tested and output them to the texture supervision module TSM. The first layer feature map is output to the target texture learning module TTM, and the highest layer feature map is output to the semantic fusion module SFM.

[0113] The Target Texture Learning Module (TTM) is used to extract the target texture feature map based on the first-layer feature map and output it to the Texture Supervision Module (TSM).

[0114] The texture supervision module TSM is used to perform deep feature extraction at each level based on the target texture feature map output by the target texture learning module TTM and the feature maps at each level output by the feature extractor, and to obtain the deep feature maps at each level and output them to the semantic fusion module SFM.

[0115] The Semantic Fusion Module (SFM) is used to obtain fused features at each level based on the highest-level feature map output by the feature extractor and the depth feature maps at each level output by the Texture Supervision Module (TSM). It then performs semantic fusion between the depth feature maps at each level and the fused features at the next higher level to output the camouflaged target detection results.

[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0117] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A camouflage target detection method based on texture-guided, characterized in that, Comprise: S1. Collect the image to be tested and input to the trained camouflage target detection model; S2. Process the image to be tested to output the camouflage target detection result; The specific content of S2 comprises: S21. Extract the feature maps of each level of the input image to be tested through the feature extractor and output to the texture supervision module, and output the first layer feature map to the target texture learning module, and output the highest layer feature map to the semantic fusion module; S22. The target texture learning module extracts the target texture feature map according to the first layer feature map and outputs it to the texture supervision module; S23. The texture supervision module performs deep feature extraction at each level according to the target texture feature map output by the target texture learning module and the feature maps of each level output by the feature extractor, obtains the deep feature maps of each level and outputs them to the semantic fusion module; S24. The semantic fusion module obtains the fusion features of each level according to the highest layer feature map output by the feature extractor and the deep feature maps of each level output by the texture supervision module, fuses the deep feature maps of each level with the fusion features of the next higher level, and outputs the camouflage target detection result; Step S22, the specific content of extracting the target texture feature map is: using 1x1 convolution to reduce the channel number of the first layer feature map to C1 / 4, using multi-scale convolution The different scale features of the first layer feature map X1 are extracted, multi-scale features are obtained by using two ConvBRU layers for integration, the ConvBRU layer includes 3x3 convolution, Batch Normalization, ReLU and bilinear interpolation up-sampling, and the feature map size is restored to the same size as the input image I to be tested, and an output target texture feature map T is obtained out ; Step S23, the specific content of deep feature extraction of each level is: performing channel dimension reduction and multi-scale convolution on each level feature map output by the feature extractor, and performing ConvBRD operation on the target texture feature map, ConvBRD is 3*3 convolution, Batch Normalization, ReLU and down-sampling, then performing element multiplication and element addition with each level feature map X i , as the deep feature map S of each level output by the texture supervision module i .

2. The texture-guided camouflage target detection method of claim 1, wherein, The calculation process of the target texture learning module S22 is: wherein, is an input image I ∈ R 3×H×W is a feature map of the first layer, is a feature map of the first layer, is a convolution operation on the input using a convolution layer with a kernel width of 2i-1 and a number of channels of Ci / 4, and Concat(·) is a channel-wise concatenation operation.

3. The texture-guided camouflage target detection method of claim 1, wherein, The calculation process of the texture supervision module S23 is: TSM(X i ,T out ) wherein, T is the target texture feature map output by the target texture learning module, and ConvBRD is 3x3 convolution, Batch Normalization, ReLU, and down-sampling. out wherein, out T is the target texture feature map output by the target texture learning module, and ConvBRD is 3x3 convolution, Batch Normalization, ReLU, and down-sampling.

4. The texture-guided camouflage target detection method of claim 1, wherein, The specific content of the semantic fusion module S24 is: S241. Receive the highest layer feature map output by the feature extractor and the deep feature maps of each level output by the texture supervision module; S242. After the highest layer feature map output by the feature extractor and the highest level deep feature map output by the texture supervision module are respectively subjected to channel dimension reduction and downsampling operations, they are subjected to channel level merging, and the highest level fusion feature and the segmentation map are output after four convolution blocks; S243. After the deep feature maps of each level output by the texture supervision module and the fusion features of the next higher level are respectively subjected to channel dimension reduction and downsampling operations, they are subjected to channel level merging to realize the fusion of different order information, and then each level fusion feature and segmentation map are output after four convolution blocks, until the first layer segmentation map is obtained as the camouflage target detection result.

5. The texture-guided camouflage target detection method of claim 1, wherein, The training content of the camouflage target detection model comprises: Pre-collect image data set containing camouflage targets and label, get image I and camouflage target label GT of each sample; Divide the pre-collected image data set containing camouflage targets into training set and test set; Train the camouflage target detection model using the training set; Test the trained camouflage target detection model using the test set.

6. The texture-guided camouflage target detection method of claim 5, wherein, The optimization target of the camouflage target detection model training is: L = L GT + L Tex Wherein, GT is the camouflage target label of the input image I, L GT is the BCE and IOU loss between the output of the semantic fusion module and GT Tex is the L1 loss between the output of the target texture learning module and the target texture map; wherein GT (w,h) and are pixels in the camouflage target label and the model output camouflage target detection result, respectively, W and H are the width and height of the image, GT (w,h) ∈ {0, 1} represents whether the pixel belongs to the camouflage target; wherein T out(w,h) is the output target texture feature map T out is the output target texture feature map T (w,h) is the output target texture feature map T 7. The texture-guided camouflage target detection method of claim 6, wherein, The method for obtaining the target texture map is: using local binary pattern LBP to extract the texture feature map of the input image, extracting the target texture feature according to the camouflage target label GT to obtain the target texture map, specifically: T=LBP(I)*GT.

8. A camouflage target detection system based on texture guide, based on the camouflage target detection method based on texture guide in any one of claims 1-7, characterized in that, Comprise an image acquisition module and a trained camouflage target detection model; The image acquisition module is used to collect the image to be tested and input to the trained camouflage target detection model; The trained camouflage target detection model is used to process the image to be tested to output the camouflage target detection result; The camouflage target detection model comprises a feature extractor, a target texture learning module, a texture supervision module and a semantic fusion module; The feature extractor is configured to extract hierarchical feature maps of the input image to be tested and output the hierarchical feature maps to the texture supervision module, output a first-layer feature map to the target texture learning module, and output a highest-layer feature map to the semantic fusion module. The target texture learning module is configured to extract a target texture feature map according to the first-layer feature map and output the target texture feature map to the texture supervision module. The texture supervision module is configured to perform deep feature extraction at each level according to the target texture feature map output by the target texture learning module and the hierarchical feature maps output by the feature extractor, obtain deep feature maps at each level, and output the deep feature maps to the semantic fusion module. The semantic fusion module is configured to obtain hierarchical fusion features according to the highest-layer feature map output by the feature extractor and the deep feature maps at each level output by the texture supervision module, perform semantic fusion between the deep feature maps at each level and the fusion features at a higher level, and output a camouflage target detection result.

Images