A camouflage target detection method based on texture and edge enhancement

Abstract

The application discloses a camouflage target detection method based on texture and edge enhancement, and belongs to the technical field of computer vision target detection. The method constructs an end-to-end detection network comprising a feature extraction backbone network, a hybrid encoder neck network and a Transform decoder detection head. By introducing a multi-scale edge enhancement module in the backbone network, the edge features of the target are enhanced, and the edge blur problem is solved. In the neck network, the texture feature enhancement and fusion of the features output by the feature extraction and edge enhancement module are carried out, multi-directional texture features are extracted, and the texture distinguishing ability is improved. Finally, the target category and position are output through the Transform decoder. The application effectively improves the detection precision and positioning accuracy of the camouflage target in a complex background, and has high practical value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision target detection technology, specifically relating to a method for detecting camouflaged targets based on texture and edge enhancement. Background Technology

[0002] Military Camouflaged People Detection (MCPD) aims to identify people or objects with camouflaged features in complex outdoor environments. Currently, object detection technology primarily relies on Convolutional Neural Networks (CNNs) and Transformer architectures. While general-purpose object detection models (such as the YOLO series, RCNN series, and RT-DETR) perform excellently in typical scenarios, they exhibit significant shortcomings when dealing with camouflaged target detection, failing to meet the demands of high-precision reconnaissance.

[0003] (1) Existing technologies struggle to extract fuzzy edge features: One of the core mechanisms of camouflage is disrupting the continuous contours of the target. Existing models (including the standard backbone network of RT-DETR) mainly focus on low-frequency semantic information when performing downsampling feature extraction, easily losing high-frequency details. This results in the edge information of the camouflaged target being severely smoothed or lost in the deep feature map, making it impossible for the model to accurately regress the bounding box of the target, leading to inaccurate localization.

[0004] (2) Existing technologies cannot effectively distinguish deceptive textures: The surface of a camouflaged target is usually covered with textures that simulate the surrounding environment (such as leaves, grass, rocks). Most existing attention mechanisms are based on global statistical information and lack the ability to analyze the frequency and directional features of textures. This makes the model easily confused by similar natural textures in the background, resulting in a large number of false detections (identifying the background as the target) or false negatives (failing to separate the target from the background).

[0005] (3) Existing models have a single feature fusion method: Existing technologies often simply superimpose or splice features during the feature fusion stage, failing to strengthen the weak features of the disguised target, resulting in the inability of deep semantics and shallow details to complement each other effectively. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to improve the ability to extract blurred edges, distinguish deceptive textures, and effectively fuse multi-scale features in the detection of camouflaged targets, so as to improve detection accuracy and positioning accuracy.

[0007] To address the aforementioned technical problems, this invention provides a camouflaged target detection method based on texture and edge enhancement. This method constructs an end-to-end detection network, which mainly consists of three parts: a feature extraction backbone network, a hybrid encoder neck network, and a Transformer decoder detection head. These components work together to solve the problems of edge blurring, texture obfuscation, and feature fusion. See below for details:

[0008] In a first aspect, the present invention provides a method for detecting camouflaged targets based on texture and edge enhancement. The method is applied to a detection network, which includes a feature extraction backbone network, a hybrid encoder neck network, and a Transformer decoder detection head. The method includes:

[0009] The input image is fed into the feature extraction backbone network;

[0010] The multi-scale edge enhancement module in the feature extraction backbone network performs multi-scale feature extraction on the input image and enhances the edge features of the target in the input image to obtain edge enhancement features. The multi-scale edge enhancement module adopts a four-way parallel branch structure.

[0011] The features output from the feature extraction backbone network are input into the hybrid encoder neck network;

[0012] The Gabor-Maximum Pooling texture enhancement module in the neck network of the hybrid encoder enhances and fuses the input features to obtain texture enhancement and fused features.

[0013] The texture enhancement and fusion features are input into the Transformer decoder detection head;

[0014] The Transformer decoder head outputs the class confidence score and bounding box coordinates of the target in the input image.

[0015] Optionally, the step of performing multi-scale feature extraction and edge feature enhancement on the input image through the multi-scale edge enhancement module in the feature extraction backbone network includes:

[0016] The input features of the multi-scale edge enhancement module are respectively input into four adaptive average pooling layers of different scales;

[0017] The features output by the four adaptive average pooling layers are upsampled to the same size using bilinear interpolation.

[0018] Each feature sampled along the path is subjected to edge enhancement processing;

[0019] The four edge-enhanced features are concatenated along the channel dimension and fused using a 1×1 convolution to obtain the edge-enhanced features.

[0020] Optionally, the edge enhancement process includes:

[0021] For the input features of the multi-scale edge enhancement module, they first pass through an average pooling layer with a kernel size of 3×3 and a stride of 1 to extract low-frequency background information;

[0022] Subtracting the low-frequency background information from the input features yields the high-frequency edge residual;

[0023] The high-frequency edge residuals are subjected to convolution with a kernel size of 3×3 and a stride of 1, and the processed result is superimposed back onto the input features to complete edge enhancement.

[0024] Optionally, the step of enhancing and fusing the input features through the Gabor-Maximum Pooling texture enhancement module in the neck network of the hybrid encoder includes:

[0025] The input features are processed by the multi-scale max pooling branch of the Gabor-Max Pooling texture enhancement module, which has kernel sizes of 3×3, 5×5, 7×7 and 9×9, respectively, and the four pooled features are then concatenated.

[0026] The stitched features are processed using a channel attention mechanism to enhance key texture channels;

[0027] The Gabor filtering branch of the Gabor-maximum pooling texture enhancement module extracts the frequency texture features of the input features in multiple preset directions using a multi-scale, multi-directional Gabor filter bank.

[0028] The features processed by the multi-scale max pooling branch and the features extracted by the Gabor filtering branch are concatenated along the channel dimension to obtain the texture enhancement fusion feature.

[0029] Optionally, the plurality of preset directions include four directions: 0°, 45°, 90°, and 135°.

[0030] Optionally, the multi-scale edge enhancement module is located in the downsampling stage of the feature extraction backbone network.

[0031] Optionally, the step of detecting the target's class confidence and bounding box coordinates through the Transformer decoder head includes:

[0032] Based on the texture enhancement fusion features, the top K high-confidence feature points are selected, and a set of learnable object query position embeddings and content embeddings are initialized using the position coordinates and semantic features of the selected feature points, respectively.

[0033] The decoder uses a multi-head self-attention mechanism and a cross-attention mechanism to enable the object query to interact with the texture enhancement fusion feature;

[0034] The object query after interaction is processed by a multilayer perceptron to generate the category confidence score and the bounding box coordinates.

[0035] Secondly, the present invention provides a camouflage target detection system based on texture and edge enhancement, comprising:

[0036] The feature extraction and edge enhancement module is used to perform multi-scale feature extraction on the input image and enhance the edge features of the target in the input image through the multi-scale edge enhancement module in the feature extraction backbone network.

[0037] The texture enhancement and fusion module is used to enhance and fuse the texture features output by the feature extraction and edge enhancement module through the Gabor-maximum pooling texture enhancement module in the hybrid encoder neck network.

[0038] The detection output module is used to predict the class confidence and bounding box coordinates of the target in the input image based on the features output by the texture enhancement and fusion module through the Transformer decoder detection head.

[0039] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the camouflage target detection method based on texture and edge enhancement as described above.

[0040] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the camouflage target detection method based on texture and edge enhancement as described above.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] 1. To address the technical problem of inaccurate localization caused by blurred edges of camouflaged targets, this invention embeds a multi-scale edge enhancement module (using four parallel branches and edge enhancement processing based on the difference principle) into the downsampling stage of the feature extraction backbone network. This effectively extracts and preserves high-frequency edge details of the camouflaged target, significantly alleviating the problem of contour information loss caused by network downsampling. Experiments show that this scheme achieves an mAP50 of 97.5% on the CPDD dataset, a 2.3% improvement over the benchmark model RT-DETR, and significantly improves the accuracy of bounding box regression.

[0043] 2. To address the technical problem of indistinguishable camouflage textures from background textures, leading to false positives and false negatives, this invention designs a Gabor-maximum pooling texture enhancement module (integrating multi-scale max pooling branches and multi-directional Gabor filtering branches) in the neck network of the hybrid encoder. This module achieves fine-grained analysis and enhancement of texture frequency and directional features. It is specifically designed to distinguish the differences in texture patterns between camouflaged targets and natural backgrounds, effectively reducing the false positive and false negative rates against complex backgrounds.

[0044] 3. To address the technical problems of existing feature fusion methods being singular and the difficulty in effectively complementing deep semantics and shallow details, this invention constructs a progressive and targeted processing flow, comprising a multi-scale edge enhancement module for early processing, a texture enhancement module for mid-level processing, and a Transformer decoder for later decision-making. This hierarchical fusion mechanism effectively strengthens and complements features at different levels, improving the overall feature representation and discrimination capabilities for camouflaged targets.

[0045] 4. Addressing the stringent real-time requirements of reconnaissance applications, this invention employs a highly efficient parallel structure for multi-scale edge enhancement modules and a streamlined design for each fusion stage. This significantly improves detection accuracy while effectively controlling model complexity and computational overhead. Experiments show that the proposed scheme has 17.7 M model parameters and 56.6 GFLOPs of computation, achieving leading detection accuracy while meeting the demands of practical applications for processing efficiency and real-time performance, thus achieving a balance between accuracy and efficiency.

[0046] In summary, this invention systematically solves the problems of edge blurring, texture confusion, and feature fusion in camouflaged target detection through the above-mentioned interrelated and synergistic technical means, thereby improving the overall detection accuracy, robustness, and practicality. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the overall model structure of the camouflage target detection method based on texture and edge enhancement provided by the present invention.

[0049] Figure 2 This is a schematic diagram of the multi-scale edge enhancement module structure provided by the present invention.

[0050] Figure 3 This is a schematic diagram of the Gabor-Maximum Pooling Texture Enhancement Module provided by the present invention.

[0051] Figure 4 This is a comparison chart of the detection performance of the TEA-DETR model provided in this embodiment of the invention and existing models on the CPDD dataset. Detailed Implementation

[0052] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] This embodiment provides a camouflage target detection method based on texture and edge enhancement, which aims to solve the technical problems of difficult edge extraction, low texture discrimination and single feature fusion method in camouflage target detection.

[0055] See Figure 1 The detection network constructed in this invention is an end-to-end deep learning model, mainly consisting of three parts: a feature extraction backbone network, a hybrid encoder neck network, and a Transformer decoder detection head, which are connected sequentially to form a complete processing flow. The feature extraction backbone network is responsible for receiving the input reconnaissance image and performing multi-scale feature extraction. The hybrid encoder neck network receives the multi-scale features output by the backbone network and performs feature enhancement and fusion. Based on the fused features, the Transformer decoder detection head finally outputs the category confidence score and precise bounding box coordinates of the camouflaged target in the image.

[0056] 1. Feature extraction backbone network and multi-scale edge enhancement module

[0057] To address the issue of easily losing edge information of camouflaged targets during downsampling in existing models, this embodiment embeds a multi-scale edge enhancement module in the downsampling stage of the feature extraction backbone network. The specific structure is shown below. Figure 2 .

[0058] This multi-scale edge enhancement module employs a four-way parallel branch structure. For the input image feature map... First, the feature maps are fed into four adaptive average pooling layers of different scales to scale them to different sizes. , , , This allows for the capture of contextual information under different receptive fields. Subsequently, bilinear interpolation is used to upsample the feature maps output by the four branches to the same size to ensure spatial resolution alignment.

[0059] Each sampled feature map needs to be processed by an edge enhancement block. The specific operation of this edge enhancement block is as follows: For the input features, firstly, a 3×3 kernel with a stride of 1 average pooling layer is used to extract low-frequency background information; then, this low-frequency background information is subtracted from the original input features to obtain the edge residual representing high-frequency details; finally, this edge residual is processed by a 3×3 kernel with a stride of 1 convolution, and the processed result is superimposed back onto the original input features, thus completing the edge feature enhancement. This process utilizes the difference principle, effectively highlighting the target contour information that is easily smoothed out in conventional convolution operations.

[0060] After edge enhancement processing of the four branches, the four output features are concatenated along the channel dimension and fused through a 1×1 convolutional layer to finally output multi-scale features with enhanced edges. This design ensures that the network can preserve and enhance the complete edge contours of the camouflaged target while extracting deep semantic information, laying the foundation for subsequent accurate localization.

[0061] 2. Hybrid encoder neck network with Gabor-maximum pooling texture enhancement module

[0062] To effectively distinguish target camouflage textures from complex background textures, this embodiment introduces a Gabor-maximum pooling texture enhancement module in the neck network of the hybrid encoder, the structure of which is shown below. Figure 3 .

[0063] This module contains two branches that work in parallel: a multi-scale max pooling branch and a Gabor filtering branch.

[0064] In the multi-scale max pooling branch, the input feature maps are fed in parallel into four max pooling layers with different receptive fields, and their pooling kernel sizes are set to 3×3, 5×5, 7×7, and 9×9, respectively. These four pooling operations aim to extract the most salient texture responses at different scales. Subsequently, the four pooled feature maps are concatenated along the channel dimension, and a channel attention mechanism is introduced.

[0065] The channel attention mechanism specifically includes: performing global average pooling on the concatenated features in the spatial dimension to obtain channel statistical descriptors; subsequently, generating weight coefficients for each channel through a dimensionality reduction mapping network containing fully connected layers and non-linear activation functions (such as ReLU), an up-dimensional mapping network with fully connected layers, and a sigmoid activation function; finally, multiplying these weight coefficients with the original concatenated features channel by channel to achieve adaptive weighting. This attention mechanism can automatically learn and assign higher weights to texture channels that are crucial for identifying camouflaged targets, thereby suppressing interference from irrelevant background textures.

[0066] In the Gabor filtering branch, a set of Gabor filter banks with preset directions is used to filter the input feature map. In a preferred embodiment, a kernel size of 31×31, a scale of 4, and an aspect ratio of [missing information] are used. A Gabor filter bank, with orientations set to [0°, 45°, 90°, 135°], is used to cover the variability of texture orientation in natural scenes. The center spatial frequency of this Gabor filter bank can be set in the range of 0.1 to 0.4 cycles / pixel, with a bandwidth of approximately one octave and a phase... Take 0 or Based on spatial spectral analysis of typical camouflage target textures, such as jungle camouflage and desert spots, and common natural background textures, presets are made to maximize the response difference between the target and the background. This embodiment uses a multi-scale Gabor filter bank to cover texture structures of different sizes. After the filtered multi-scale, multi-directional texture features are scale-aligned using methods such as bilinear interpolation, they are concatenated or element-wise added with the original input features or convolutional features in the channel dimension to achieve feature fusion. For each pixel position in the feature map, its filtered response in each direction is calculated using a coordinate rotation formula, thereby extracting frequency texture features with direction selectivity.

[0067] Finally, the salient texture features output from the multi-scale max-pooling branch and the directional texture features output from the Gabor filtering branch are concatenated along the channel dimension to obtain a fused feature map with comprehensively enhanced texture information. This module significantly improves the model's ability to identify deceptive textures by combining scale and orientation analysis.

[0068] 3. Transformer decoder detection head

[0069] The detection head in this embodiment employs a Transformer-based decoder architecture. First, based on the texture enhancement fusion feature map output by the hybrid encoder neck network, a lightweight prediction branch calculates the confidence distribution of each spatial location belonging to the foreground target in the feature map. This branch is optimized along with the main detection network during training. In one specific implementation of this invention, the lightweight prediction branch consists of a 1×1 convolutional layer with the same number of input channels as the texture enhancement fusion feature map, and the number of output channels corresponding to the total number of target categories to be detected (e.g., in camouflage target detection, it could be "camouflaged personnel" and "background"). This convolutional layer is directly followed by a Softmax or Sigmoid activation function to map the output to the probability distribution of each spatial location belonging to each category. This design ensures effective generation of foreground confidence while minimizing the computational complexity and number of parameters of the branch, meeting the overall network's efficiency requirements.

[0070] Subsequently, based on the confidence distribution, the top K spatial locations with the highest confidence ranking are selected as high-confidence feature points. Here, K is a preset hyperparameter. The value of K needs to be balanced between detection recall and decoder computational efficiency: a K value that is too small may lead to missed detection of real targets, reducing recall; a K value that is too large will increase the computational overhead of the attention mechanism in the subsequent Transformer decoder, affecting detection speed. Based on the characteristics of the camouflaged target detection task targeted by this invention, and the network structure and input resolution used, experiments have verified that setting K to 300 can maintain a high recall rate (reaching 95.4% as shown in Table 1) on typical camouflaged datasets such as CPDD while keeping the total computational cost of the model at a low level (56.6 GFLOPs). Therefore, K=300 is a preferred parameter configuration in this embodiment.

[0071] Next, using the location coordinates and semantic features of these selected feature points, a set of learnable object query location embeddings and content embeddings are initialized respectively.

[0072] Subsequently, the decoder enables these object queries to interact through a multi-head self-attention mechanism, and enables each object query to interact with the texture enhancement fusion feature map through a cross-attention mechanism, thereby encoding image feature information into the query vector.

[0073] Finally, each object query vector after interaction is fed into a multilayer perceptron (MLP), which outputs two prediction results in parallel: a confidence score of the target's category and a four-dimensional vector representing the coordinates of the four corners of the predicted bounding box, thereby completing the target identification and localization.

[0074] To verify the effectiveness of the method of the present invention, the following experiments were conducted in this embodiment:

[0075] The method in this embodiment (hereinafter referred to as TEA-DETR) is implemented based on Python 3.8 and the PyTorch 1.10.1 deep learning framework, using Ubuntu 20.04 LTS as the operating system and a high-performance workstation equipped with an NVIDIA GeForce RTX4090D graphics card (24GB of VRAM) and an Intel Core i9 processor. The publicly available CPDD dataset is used for training and testing, containing 2600 images covering scenes such as woodland, wasteland, and snowfields. The training and testing sets are randomly divided in an 8:2 ratio.

[0076] During training, the input images were uniformly scaled to 480×854 pixels. The AdamW optimizer was used with an initial learning rate of 0.1 and a learning rate decay strategy, ultimately reducing the learning rate to 0.0001. The batch size was set to 8, and training lasted for 300 epochs.

[0077] During the training phase, this invention employs a combined loss function for optimization. First, the algorithm establishes a one-to-one correspondence between predicted and ground truth boxes. Then, the following loss terms are calculated for the output of each layer of the Transformer decoder:

[0078] (1) Classification loss For category prediction of supervised object queries, this embodiment preferably uses Focal-type loss such as Varifocal Loss;

[0079] (2) Bounding box regression loss The L1 loss function is used to measure the difference between the predicted bounding box and the ground truth bounding box in terms of center point coordinates, width, and height.

[0080] (3) Generalized intersection-union (GIoU) ​​loss This is used to further optimize the spatial overlap between the predicted bounding box and the ground truth bounding box.

[0081] Total loss The weighted sum of the above losses, i.e. In one specific embodiment, the loss weights can be set as follows: classification loss weights Bounding box L1 loss weights GIoU loss weights These weight parameters can be adjusted according to the characteristics of the dataset for the actual task. The configuration in this embodiment achieved a good balance between detection accuracy and training stability in the experiments.

[0082] This embodiment uses mean accuracy (mAP), precision, and recall as the main evaluation metrics, while also calculating the number of model parameters and computational cost to assess efficiency. TEA-DETR is compared with current mainstream models such as RT-DETR, YOLOv8m, and YOLOv10m on the CPDD test set.

[0083] Experimental results are as follows Figure 4 As shown, the comparison of detection performance of different models in the same camouflaged scene is presented intuitively. The quantitative results are shown in the table below:

[0084] Table 1 Comparison of test results for different models

[0085]

[0086] Experimental results show that the TEA-DETR model of this invention achieves significantly better performance than existing technologies on the CPDD dataset. The mAP50 (mean accuracy at an intersection-over-union ratio of 0.5) reaches 97.5%, a 2.3% improvement over the benchmark model RT-DETR. The mAP95 (mean accuracy at an intersection-over-union ratio of 0.95) reaches 62.1%, both superior to the YOLOv10m and RT-DETR models. Furthermore, the model of this invention has only 17.7 M parameters and a computational cost of 56.6 GFLOPs, both of which are better than RT-DETR.

[0087] Example 2

[0088] This embodiment provides a camouflage target detection system based on texture and edge enhancement. This system is used to implement the camouflage target detection method based on texture and edge enhancement described in Embodiment 1, including:

[0089] The feature extraction and edge enhancement module is used to perform multi-scale feature extraction on the input image and enhance the edge features of the target in the input image through the multi-scale edge enhancement module in the feature extraction backbone network;

[0090] The texture enhancement and fusion module is used to enhance and fuse the texture features output by the feature extraction and edge enhancement modules by using the Gabor-maximum pooling texture enhancement module in the hybrid encoder neck network.

[0091] The detection output module is used to predict the class confidence and bounding box coordinates of targets in the input image based on the features output by the texture enhancement and fusion module, using the Transformer decoder detection head.

[0092] Example 3

[0093] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the camouflage target detection method based on texture and edge enhancement as described in Embodiment 1.

[0094] Example 4

[0095] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the camouflage target detection method based on texture and edge enhancement as described in Embodiment 1.

[0096] In summary, this invention forms a complete solution for camouflaged target detection through the synergistic effect of the multi-scale edge enhancement module in the feature extraction backbone network, the Gabor-maximum pooling texture enhancement module in the hybrid encoder neck network, and the Transformer decoder detection head. This solution effectively solves the technical challenges of edge blurring and texture confusion, significantly improving detection accuracy and robustness in complex camouflaged environments while maintaining the model's inference efficiency.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting camouflaged targets based on texture and edge enhancement, characterized in that, The method is applied to a detection network, which includes a feature extraction backbone network, a hybrid encoder neck network, and a Transformer decoder detection head; the method includes: The input image is fed into the feature extraction backbone network; The multi-scale edge enhancement module in the feature extraction backbone network performs multi-scale feature extraction on the input image and enhances the edge features of the target in the input image to obtain edge enhancement features. The multi-scale edge enhancement module adopts a four-way parallel branch structure. The features output from the feature extraction backbone network are input into the hybrid encoder neck network; The Gabor-Maximum Pooling texture enhancement module in the neck network of the hybrid encoder enhances and fuses the input features to obtain texture enhancement and fused features. The texture enhancement and fusion features are input into the Transformer decoder detection head; The Transformer decoder head outputs the class confidence and bounding box coordinates of the target in the input image. The step of extracting and enhancing edge features of the input image through the multi-scale edge enhancement module in the feature extraction backbone network includes: The input features of the multi-scale edge enhancement module are respectively input into four adaptive average pooling layers of different scales; The features output by the four adaptive average pooling layers are upsampled to the same size using bilinear interpolation. Each feature sampled along the path is subjected to edge enhancement processing; The four edge-enhanced features are concatenated along the channel dimension and fused using a 1×1 convolution to obtain the edge-enhanced features. The edge enhancement process includes: For the input features of the multi-scale edge enhancement module, they first pass through an average pooling layer with a kernel size of 3×3 and a stride of 1 to extract low-frequency background information; Subtracting the low-frequency background information from the input features yields the high-frequency edge residual; The high-frequency edge residuals are subjected to convolution with a kernel size of 3×3 and a stride of 1, and the processed result is superimposed back onto the input features to complete edge enhancement. The Gabor-Maximum Pooling texture enhancement module in the neck network of the hybrid encoder enhances and fuses the input features, including: The input features are processed by the multi-scale max pooling branch of the Gabor-Max Pooling texture enhancement module, which has kernel sizes of 3×3, 5×5, 7×7 and 9×9, respectively, and the four pooled features are then concatenated. The stitched features are processed using a channel attention mechanism to enhance key texture channels; The Gabor filtering branch of the Gabor-maximum pooling texture enhancement module extracts the frequency texture features of the input features in multiple preset directions using a multi-scale, multi-directional Gabor filter bank. The features processed by the multi-scale max pooling branch are concatenated with the features extracted by the Gabor filtering branch by channel dimension to obtain the texture enhancement fusion feature. The multiple preset directions include four directions: 0°, 45°, 90°, and 135°. The process of detecting the target's class confidence and bounding box coordinates through the Transformer decoder head includes: Based on the texture enhancement fusion features, the top K high-confidence feature points are selected, and a set of learnable object query position embeddings and content embeddings are initialized using the position coordinates and semantic features of the selected feature points, respectively. The decoder uses a multi-head self-attention mechanism and a cross-attention mechanism to enable the object query to interact with the texture enhancement fusion feature; The object query after interaction is processed by a multilayer perceptron to generate the category confidence score and the bounding box coordinates.

2. The camouflage target detection method based on texture and edge enhancement according to claim 1, characterized in that, The multi-scale edge enhancement module is located in the downsampling stage of the feature extraction backbone network.

3. A camouflaged target detection system based on texture and edge enhancement, characterized in that, The system for implementing the camouflage target detection method based on texture and edge enhancement as described in any one of claims 1 to 2, the system comprising: The feature extraction and edge enhancement module is used to perform multi-scale feature extraction on the input image and enhance the edge features of the target in the input image through the multi-scale edge enhancement module in the feature extraction backbone network. The texture enhancement and fusion module is used to enhance and fuse the texture features output by the feature extraction and edge enhancement module through the Gabor-maximum pooling texture enhancement module in the hybrid encoder neck network. The detection output module is used to predict the class confidence and bounding box coordinates of the target in the input image based on the features output by the texture enhancement and fusion module, using the Transformer decoder detection head.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the camouflage target detection method based on texture and edge enhancement as described in any one of claims 1 to 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the camouflage target detection method based on texture and edge enhancement as described in any one of claims 1 to 2.

Images