A camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition

Abstract

This invention relates to the field of computer vision and target detection technology, specifically a camouflaged target detection method based on dynamic Top-k feature selection and cross-tensor decomposition. The method includes: inputting a camouflaged target image into an encoder network to extract multi-level features, generating feature maps at different scales; inputting these features into a global perception module and a local optimization module for feature extraction at different scales, outputting global and local features; performing cross-tensor decomposition on the global and local features, generating complementary features through low-rank factor matrix expansion and cross-merging; inputting the complementary global and local features, along with the previous layer's fused features, into a hierarchical fusion module; inputting the features into a hybrid weighted decoder, performing inverse optimization using the channel attention mechanism of the DSE module, and finally outputting the camouflaged target detection result; and performing multi-level supervision on the output segmentation map and the final prediction map, jointly optimizing model parameters through a loss function to flexibly adjust the dependence on different features.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, specifically a method for detecting camouflaged targets based on dynamic Top-k feature selection and cross tensor decomposition. Background Technology

[0002] In nature, many animals are adept at camouflage, such as the dead leaf butterfly and the stick insect. They blend into the environment to better avoid predators. Animal camouflage typically involves altering its color and shape to blend into the background, making it difficult for predators to spot; or using special patterns to create false edges, making its true outline less visible. Artificial camouflage shares similarities with animals, such as using body paint and camouflage clothing to conceal oneself.

[0003] Early traditional camouflaged object detection (COD) methods typically relied on handcrafted features or low-level visual priors (e.g., color, texture, and intensity) to detect camouflaged objects. However, these features or priors may not fully capture the complex structure and content within an image. In particular, objects occluded by cams in natural images often have extremely high similarity to the background, leading to unsatisfactory segmentation results from traditional strategies.

[0004] Later, due to the availability of large-scale datasets, many deep learning-based COD methods were proposed to solve this difficult segmentation problem. Most of these methods aim to aggregate multi-scale features captured by various convolutional operations to gradually distinguish objects from the background. They achieve better performance by integrating spatial information containing different receptive fields. However, the receptive field of convolutional operations is often limited, and the extracted cues are often local, making it difficult to establish a global relationship model between all pixels. Especially when the model's receptive field is limited to a local perspective and lacks a global perspective, the predicted camouflaged object may be an incomplete whole. Furthermore, camouflaged objects in images are often varied in shape and scale, requiring a large amount of contextual information to understand the image content. Therefore, modeling global contextual information and capturing long-distance dependencies between pixels is particularly important in camouflaged object detection. Traditional convolutional operations may lose some key details during feature extraction, and the information transfer efficiency between different levels of features is limited, making it difficult to fully utilize the correlation between features. Moreover, in the subsequent fusion process, redundant learning of features between different levels often leads to inaccurate and inefficient feature fusion. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a camouflaged target detection method based on dynamic Top-k feature selection and cross-tensor decomposition. It extracts global features by modeling the overall contextual relationships of the camouflaged image using Mamba blocks in the global perception module. The linear complexity of the global Mamba's long-sequence modeling efficiently captures semantic consistency across regions. The Mamba blocks in the local optimization module focus on local details, avoiding the loss of fine-grained information caused by global pooling or downsampling. The local Mamba, while preserving high-resolution feature maps, can model and enhance the expressive power of fine-grained features. By introducing a dynamic Top-k feature selection mechanism in both the global and local Mambas, key features are dynamically filtered to remove background features irrelevant to the target, enhancing the target region response and suppressing irrelevant background noise. Cross-tensor decomposition of global and local features effectively reduces the learning interference of redundant feature information between different levels during subsequent fusion. Processing and cross-fusion of global, local, and previous-layer fusion features in the hierarchical fusion module more comprehensively captures the relationships between features of different levels and types, enhancing feature representation capabilities. Adopting different feature interaction paths also avoids the loss of important information. In the hybrid weighted decoder, different information is integrated through cross-layer aggregation, dynamic weighting and inverse optimization. This allows for the mining and utilization of potentially important information with different features. The decoder introduces a dynamic weighting mechanism, which enables features at different levels to be adaptively weighted according to their importance during the fusion process, flexibly adjusting the degree of dependence on different features.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition, characterized in that: the method includes:

[0008] S1, Disguise the target image I∈R H×W×C Input encoder network extracts multi-level features E i,i＝1,2,3,4,5 Generate resolution of Feature maps at different scales;

[0009] S2. Input the features of layers E2-E5 into the global perception module and the local optimization module respectively to extract features at different scales, filter out the global key features and local sensitive features, and output the global features G2-G5 and the local features L2-L5.

[0010] S3, Global features G at the same level i and local features L i Perform cross tensor decomposition, and generate complementary features through low-rank factor matrix expansion and cross merging;

[0011] S4. Combine the complementary global features, local features, and the previous layer's fused features F.i-1 The input layer fusion module achieves cross-layer feature interaction through gated convolution and dynamic weighting mechanism, and outputs fused features F2-F5;

[0012] S5. Input the F2-F5 features into the hybrid weighted decoder, combine the channel attention mechanism of the DSE module for inverse optimization, generate and upsample the prediction map layer by layer, and finally output the camouflage target detection result. The input of the latter two decoders not only includes the fused features of the current level, but also the feature maps output by the previous two decoders.

[0013] S6. Perform multi-level supervision on the segmentation map and the final prediction map output by the hybrid weighted decoder, and jointly optimize the model parameters through the loss function.

[0014] Preferably, the encoder backbone network in step S1 uses ResNet-50 to extract five initial features. Since the resolution of the first layer feature E1 is greater than the resolution threshold and there is background noise, E1 is discarded and E2-E5 are used to complete the subsequent COD task.

[0015] Preferably, step S2 incorporates two modules to extract local and global feature information. The local refinement module enhances the spatial local information in the initial features, comprising multiple Local Visual Mamba Blocks (LVMBs). Within the LVMBs, a dynamic Top-k feature selection mechanism is introduced to dynamically filter key features from the local features, thereby filtering out irrelevant background features and improving the model's sensitivity to key regions. The LVMB module outputs local features L2-L5. The global perception module also comprises multiple Global Visual Mamba Blocks (GVMBs), utilizing GVMBs to acquire all pixel relationships from a global perspective. Similar to LVMB, the GVMB also incorporates a dynamic Top-k feature selection mechanism to filter the most critical global features for the detection task, reducing redundant feature processing. The GVMB module outputs global features G2-G5.

[0016] Preferably, Local Top-k Feature Selection (LTTS) and Global Top-k Feature Selection (GTTS) fully utilize sparsity by selecting the top k tags with the highest relevance to the query, thereby capturing the most crucial disguised target information for extraction. k is dynamically set to a series of values ​​satisfying ki = (i+1) / (n+2), where i is the current level index, n is the total number of levels, and the difference rate of k values ​​between different levels is no less than 15%. For example, if k = 1 / 2, only the top 50% of elements with the highest scores will be retained and activated, while the remaining 50% of elements will be masked as 0.

[0017] For local visual mamba blocks, an LTTS module is introduced as the feature extractor. Compared to general visual mamba block feature extractors, local visual mamba blocks include an unfolding operator, which can effectively guarantee the spatial proximity of adjacent tags in 2D or 3D arrays. A kernel size commonly used in convolutional layers is employed, with a window size R = 3. Since the number of channels in LTTS becomes C', a linear layer is inserted after the State Space Module (SSM) to project the tags onto the original C-dimensional space. For global visual mamba blocks, a GTTS module is introduced as the feature extractor. After passing through a Deep Convolutional Layer (DWC), the SiLU activation function, and flattening, it is combined with a Top-k feature selection mechanism to generate GRF-based features, which are then forwarded to the next SSM. The number of channels in the output remains unchanged, but the sequence length increases, so the linear layer after the SSM is removed.

[0018] Preferably, in step S3, the same level features L i and G i By decomposing the convolutional layers, selectively integrating local information into global information to form complementary global features, and selectively integrating global information into local information to form complementary local features, the problem of redundant feature learning between different levels in camouflage target detection is solved. For the global feature G... i and local features L i Construct low-rank factor matrices A, a∈R respectively. TS×r and Then, the dimensions of the low-rank factor matrices are respectively extended to A', a'∈R. TS×(r+Δr) and Where Δr = 16, S and T represent the number of input channels and the number of kernels learned by the convolutional layer, respectively, D1 and D2 represent the width and height of the convolutional kernel space window, respectively, and r represents the rank of the original weight matrix before decomposition; finally, the weight matrices are cross-merged to generate the enhanced weight matrices M' = A' × b' and m' = a' × B', where the parameter increment ΔP_fac = Δr(TS + D1D2), and the convolutional filters K,k can be derived from M',m'.

[0019] Preferably, in order to explicitly capture L i and G i Complementary cues between layers: We maximize the L2 or L1 distance between feature maps output from each branch of a layer and use it as an additional term in the task training objective. This can be achieved using the following loss Lc: Lc = -||K*x G -ΔK*x G -k*x L -Δk*x L || p ;

[0020] Where p = {1, 2}, * denotes convolution, and x denotes a vector. K, k and ΔK, Δk have the same dimension.

[0021] Preferably, the hierarchical fusion module in step S4 includes three branches, which can adaptively fuse local features, global representation, and semantic information from the previous layer based on the input features. 1. Global weight generation path: GlobalAvgPool → 1x1 convolution → BN → ReLU → 1x1 convolution → Sigmoid; 2. Local weight generation path: a concatenated structure of two 1×1 convolutional layers, along with batch normalization (BN) and the GELU loss function; 3. Previous-level feature weight generation path: a concatenated structure of two 1×1 convolutional layers, along with batch normalization and the ReLU loss function. The outputs of the three paths are concatenated, and then gated convolution is used to filter redundant information and perform residual connections to improve feature discrimination capability.

[0022] Furthermore, step S5 proposes a hybrid weighted decoder that integrates different information through cross-layer aggregation, dynamic weighting, and inverse optimization, enabling the mining and utilization of potentially important information with different features. The previously fused features and the segmentation features output from adjacent decoders are concatenated, convolved, reshaped, and inverted, and then added and multiplied to obtain a feature map. This resulting feature map is then added to the previous segmentation feature map after dynamic weight allocation by the DSE module, ultimately outputting a segmentation map D. i .

[0023] The DSE module of the hybrid weighted decoder includes a dual attention branch and a weight fusion part. Compared with the traditional SE module, the strong channel attention of DSE enables the model to capture the dependencies between feature channels from both average pooling and max pooling perspectives, while suppressing irrelevant channels. A dynamic weighting mechanism is introduced into the decoder, allowing features at different levels to be adaptively weighted according to their importance during the fusion process, flexibly adjusting the degree of dependence on different features.

[0024] Preferably, in step S6, the segmentation map D6 is replaced by the output of the previous stage fusion module, and the output of each subsequent decoder is also an input to the next stage decoder. The camouflaged target segmentation map D6 is generated by concatenating D6 with the four decoders. i,i＝2,3,4,5,6 The loss is calculated using the ground truth graph (GT) of the training set, and the selected structured loss function is the weighted binary cross-entropy loss. Weighted intersection and combination loss And tensor loss Lc. To handle multi-scale outputs, during training, all outputs D are processed. k＝2,3,4,5,6 Upsampling is performed to match the resolution of the ground truth segment (GT). The formula for the total loss function L of the camouflaged target detection model is as follows:

[0025]

[0026] Where k represents different network layers; D i This represents the segmentation map of camouflaged targets at different layers; GT represents the ground truth label of the camouflaged target.

[0027] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0028] This invention employs a global Mamba algorithm with linear complexity to model the overall contextual relationships of camouflaged images in the global perception module, efficiently capturing cross-regional semantic consistency. In the local optimization module, a local Mamba algorithm focuses on local details, capturing local anomalies through fine-grained contrast, preserving high-resolution feature maps while enhancing fine-grained feature representation. A dynamic Top-k feature selection mechanism is introduced into both the global and local Mamba algorithms to dynamically filter key features, removing irrelevant background features, enhancing target region response, and suppressing irrelevant background noise. Furthermore, the cross-tensor decomposition of global and local features effectively reduces the complexity of subsequent fusion processes. The learning interference of redundant feature information between the same level; in the hierarchical fusion module, global, local and previous layer fusion features are processed and cross-fused to more comprehensively capture the relationship between features of different levels and types, and enhance feature representation capabilities; in the hybrid weighted decoder, different information is integrated through cross-layer aggregation, dynamic weighting and inverse optimization, which can mine and utilize potential important information with different features. The introduction of a dynamic weight mechanism in the decoder enables features of different levels to be adaptively weighted according to their importance during the fusion process, flexibly adjusting the degree of dependence on different features, enhancing the adaptability and robustness of the model, and improving the accuracy of camouflaged target detection. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0030] Figure 1 This is a flowchart illustrating the overall implementation of the camouflage target detection method based on dynamic Top-k feature selection and cross tensor decomposition of the present invention.

[0031] Figure 2 This is a flowchart illustrating the implementation of the Local Visual Mamba Block (LVMB) according to an embodiment of the present invention.

[0032] Figure 3 This is a flowchart illustrating the Global Visual Mamba (GVMB) implementation process according to an embodiment of the present invention.

[0033] Figure 4 This is a flowchart illustrating the implementation of the cross-tensor decomposition convolutional layer according to an embodiment of the present invention.

[0034] Figure 5This is a flowchart illustrating the implementation of the hierarchical fusion module according to an embodiment of the present invention.

[0035] Figure 6 This is a flowchart illustrating the implementation of the hybrid weighted decoder according to an embodiment of the present invention;

[0036] Figure 7 This is a flowchart illustrating the implementation of Dynamic Compression and Excitation (DSE) according to an embodiment of the present invention.

[0037] Figure 8 This is a schematic diagram illustrating a qualitative comparison between six mainstream camouflage target segmentation models and the method of this invention in a comparative experiment of embodiments of the present invention. Detailed Implementation

[0038] 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.

[0039] Please see Figures 1-8 The present invention provides the following technical solution:

[0040] Example 1:

[0041] Figure 1 This is a flowchart illustrating the overall implementation of a camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition according to the present invention.

[0042] like Figure 1 As shown, the camouflage target detection network model of the present invention includes an encoder, a global perception module, a local optimization module, a cross tensor decomposition, a hierarchical fusion module, and a hybrid weighted decoder.

[0043] The backbone network uses the RestNet-50 model from the ResNet series to extract multi-scale features E of images containing camouflaged targets. i,i＝1,2,3,4,5 Since most camouflage target detection algorithms in recent years have chosen ResNet-50, ResNet-50 was also chosen as the backbone network here in order to verify the effectiveness of the module through comparative experiments.

[0044] The global perception module consists of multiple global visual mamba blocks, arranged from high to low. The outputs of adjacent mamba blocks are concatenated to output the global feature G. i,i＝2,3,4,5 The linear complexity of long sequence modeling in the Global Mamba can efficiently capture semantic consistency across regions.

[0045] The local optimization module consists of multiple local visual mamba blocks, connected in the same way as the global feature extraction structure, and outputs local features L. i,i＝2,3,4,5 The Local Mamba technique focuses on local details, preserving high-resolution feature maps while enhancing fine-grained feature representation. Simultaneously, a dynamic Top-k feature selection mechanism is introduced during the Mamba block feature extraction stage to dynamically filter key features, discard irrelevant background features, enhance the target region response, and suppress irrelevant background noise. Then, the output global feature G... i and local features L i Cross-tensor decomposition effectively reduces the learning interference of redundant feature information between different levels during subsequent fusion. The hierarchical fusion module includes global, local, and previous layer fusion features F. i+1,i＝2,3,4 The three branches and four fusion modules are adjacent and interconnected, which can more comprehensively capture the relationships between features at different levels and types, and enhance the feature representation capability.

[0046] The hybrid weighted decoder will obtain the feature F i,i＝2,3,4,5 Decode and generate prediction map D i,i＝2,3,4,5 D6 is replaced by the output of the first-level fusion module. A dynamic weighting mechanism is introduced in the decoder through the DSE module, allowing features from different levels to be adaptively weighted according to their importance during the fusion process, flexibly adjusting the weighting of different feature segmentation maps D. i+1 D i+2,i＝2,3,4 The degree of dependence. The hybrid weighted decoder integrates different information through cross-layer aggregation, dynamic weighting and inverse optimization, which can mine and utilize potentially important information with different features to obtain a more refined camouflaged target segmentation map.

[0047] The camouflaged target detection method of the present invention includes the following:

[0048] Step 1: Disguise the target image I∈R H×W×C Input encoder network extracts multi-level features E i,i＝1,2,3,4,5 Generate resolution of Feature maps at different scales;

[0049] In this implementation, the encoder backbone network uses ResNet-50 to extract five initial features. Because the first layer feature E1 has a high resolution and contains background noise, E1 is discarded, and Ei is used instead. i,i＝2,3,4,5 Complete the subsequent COD missions.

[0050] Step two, E i,i＝2,3,4,5 The features are input into the global perception module and the local optimization module respectively for feature extraction at different scales. The key global features and the locally sensitive features are selected, and the global feature G is output. i,i＝2,3,4,5 and local features L i,i＝2,3,4,5 ;

[0051] In this implementation, two modules were designed to extract local and global feature information. A local refinement module was used to enhance the spatial local information in the initial features, which consisted of multiple Local Visual Mamba Blocks (LVMBs). Within the LVMBs, a dynamic Top-k feature selection mechanism was introduced to dynamically filter key features from the local features, thereby filtering out irrelevant background features and improving the model's sensitivity to key regions. The LVMB module outputs the local feature L. i,i＝2,3,4,5 The global perception module also comprises multiple interconnected Global Visual Mamba Blocks (GVMBs), utilizing GVMBs to acquire the relationships between all pixels from a global perspective. Similar to LVMB, a dynamic Top-k feature selection mechanism is introduced in GVMBs to filter the most critical global features for the detection task, reducing redundant feature processing. The global feature G is output after passing through the GVMB module. i,i＝2,3,4,5 .

[0052] Figure 2 This is a flowchart illustrating the implementation of the Local Visual Mamba Block (LVMB) according to an embodiment of the present invention.

[0053] like Figure 2 As shown, Local Top-k Feature Selection (LTTS) fully utilizes sparsity by selecting the top k tags that are most relevant to the query, thereby capturing the most critical camouflaged target information for extraction.

[0054] Specifically, we first use DWC to compress the input channel C by a factor S, thus avoiding the computational overhead of subsequent processors. After passing the compressed features through SiLU, we unfold them using a fixed convolution kernel of size R×R. This operation allows us to copy the features and preserve the spatial relationships of nearby features. Then, we generate a dense matrix through a self-rotated dot product operation. k is dynamically set to a series of values. For example, with k1 = 1 / 2, only the top 50% of elements are retained active, while the remaining 50% are masked at 0; similarly, when k4 = 4 / 5, the sparsity is 20%. In contrast to the lack of flexibility in exploring the potential extent of sparsity by fixing the k value, the proposed dynamic selection allows for a selection process from sparse to dense by setting k to multiple values. Finally, we reshape and flatten the selected spatial dimension to form a one-dimensional sequence of local features, ensuring that adjacent tokens within the local window are all along the channel axis of the query token. The number of output channels is C' = CR. 2 / S. For local visual mamba blocks, an LTTS module is introduced as the feature extractor. Compared with the general visual mamba block feature extractor, the local visual mamba block includes an unfolding operator, which can effectively guarantee the spatial proximity of adjacent tags in the 2D or 3D array. Using the kernel size commonly used in convolutional layers, the window size R = 3 is set. Since the number of channels of LTTS becomes C', a linear layer is inserted after the state space module (SSM) to project the tags onto the original C-dimensional space.

[0055] Figure 3 This is a flowchart illustrating the Global Visual Mamba (GVMB) implementation process according to an embodiment of the present invention.

[0056] like Figure 3 As shown, Global Top-k Feature Selection (GTTS) generates global, spatially independent channel labels.

[0057] Specifically, given an input feature map of size H×W×C′, we spatially compress it using a DWC expansion with a stride of K×K, and then flatten its spatial dimensions to generate a shape of C′×HW / K. 2 The global marker is then used, and the channel size and spatial size are transposed. Similar to LTTS, a dense matrix M∈R is then generated through a dot product operation of the self-transpose. C'×C' k is dynamically set to a series of values. To improve computational efficiency, each input channel is compressed into a global label across all spatial dimensions. These steps serve to learn an approximation of the global context in each input channel, thus mitigating the loss of fine-grained details associated with expanded DWC. Finally, we project these selected labels into a C′-dimensional space using a linear layer and apply the SiLU activation function. For the global visual mamba block, a GTTS module is introduced as a feature extractor, which, after passing through a deep convolutional layer (DWC), the SiLU activation function, and flattening, is combined with a Top-k feature selection mechanism to generate features with a GRF, which are then forwarded to the next SSM. Since the number of channels in the output remains unchanged while the sequence length increases, the linear layer after the SSM is eliminated.

[0058] Step 3, combine the global features G at the same level i and local features L i Perform cross tensor decomposition, and generate complementary features through low-rank factor matrix expansion and cross-merging;

[0059] In this example, the same level feature L i and G i By decomposing the convolutional layers, selectively incorporating local information into global information to form complementary global features, and selectively incorporating global information into local information to form complementary local features, the problem of redundant feature learning between different levels in the process of camouflaged target detection is solved.

[0060] Figure 4 This is a flowchart illustrating the implementation of the cross-tensor decomposition convolutional layer according to an embodiment of the present invention.

[0061] like Figure 4 As shown, we first examine G respectively. i and L i Learn two trainable low-rank factor matrices and generate the convolutional weights for each layer through their product. The weights of the convolutional layers are... Let P = TSD ... TS×R and Initially, r is their inner dimension, representing the rank of the original weight matrix (before decomposition). To better fuse features from different levels, we enhance the network's capabilities by slightly increasing the number of trainable parameters. This is achieved by adding columns / rows to the factor matrix. The factor matrix emerging from the added columns / rows effectively acts as parallel trainable branches, enabling the network to extract key features from different levels to form complementary information. We increase r of matrices A, a, B, and b by Δr (where Δr > 0), thus producing A', a' ∈ R. TS×(r+Δr) and r+Δr serves as their new internal dimension. Where Δr = 16, and finally, they are cross-merged to generate enhanced weight matrices M' = A'×b' = M+Δm and m' = a'×B' = m+ΔM, where Δm = ΔAΔb, ΔM = ΔaΔB. The parameter increment ΔP_fac = Δr(TS + D1D2) can be derived from Δm, and the convolution filter K,k can be derived from M',m'.

[0062] In order to clearly capture L i and G i Complementary cues between layers: We maximize the L2 or L1 distance between feature maps output from each branch of a layer and use it as an additional term in the task training objective. This can be achieved using the following loss Lc: Lc = -||K*x G -ΔK*x G -k*x L -Δk*x L || p ;

[0063] Where p = {1, 2}, * denotes convolution, and x represents a matrix vector. K, k and ΔK, Δk have the same dimensions.

[0064] Step 4: Complementary global features, local features, and the previous layer's fused features F i-1 The input layer fusion module achieves cross-layer feature interaction through gated convolution and dynamic weighting mechanism, and outputs fused feature F. i,i＝2,3,4,5 ;

[0065] In this embodiment, the hierarchical fusion module comprises three branches, which can adaptively fuse local features, global representation, and semantic information from the previous layer based on the input features. These include a global weight generation path, a local weight generation path, and a previous-level feature weight generation path. The outputs of these three paths are concatenated, and then redundant information is filtered through gated convolution, followed by residual connections. Weighted fusion of features at different levels allows the model to adaptively adjust the fusion strategy according to the characteristics of the input features, improving the fusion effect. Feature interactions between different paths also prevent information loss.

[0066] Figure 5 This is a flowchart illustrating the implementation of the hierarchical fusion module according to an embodiment of the present invention.

[0067] like Figure 5 As shown, the input global feature G i First, the feature undergoes Global AvgPooling, then sequentially passes through Point-wise Convolution, Batch Normalization (BN), ReLU activation, and again through Point-wise Convolution and BN. Finally, a sigmoid function is used to obtain a weight value for subsequent feature weighting operations. Local feature Li undergoes Point-wise Convolution, BN, ReLU, again through Point-wise Convolution and BN, and then a sigmoid function to obtain another weight value. The previous layer feature F... i-1 First, Point-wise Conv is performed, followed by Global AvgPooling. Then, the intermediate result from the processed global feature branch is concatenated, and further processed by LN (Layer Normalization), Point-wise Conv, and GELU activation to obtain an intermediate feature representation. The weights obtained by applying Sigmoid to the global and local features are multiplied by the intermediate feature from the previous layer, and then these two weighted features are concatenated. The concatenated feature is then processed by GConv (Gated Convolution), multiplied and added by the previously concatenated features, and finally, the fused feature Fi is output through Point-wise Conv. Gated convolution is used to filter redundant information, improving the feature discrimination ability, and residual connections are used to avoid gradient vanishing, exploding, and network degradation to some extent, thus effectively capturing global and local feature information from various layers.

[0068] Step 5: Input the F2-F5 features into the hybrid weighted decoder, and perform inverse optimization by combining the channel attention mechanism of the DSE module. Generate and upsample the prediction map layer by layer, and finally output the camouflage target detection result. The input of the last two decoders not only includes the fused features of the current level, but also the feature maps output by the first two decoders.

[0069] In this embodiment, the hybrid weighted decoder integrates different information through cross-layer aggregation, dynamic weighting, and inverse optimization, enabling the mining and utilization of potentially important information with different features. The previously fused features and the segmentation features output from adjacent decoders are concatenated, convolved, reshaped, and inverted. The resulting feature map is then added and multiplied with the previous segmentation feature map after dynamic weight allocation by the DSE module, ultimately outputting a segmentation map D. i The DSE module of the hybrid weighted decoder includes a dual attention branch and a weight fusion part. Compared with the traditional SE module, the strong channel attention of DSE enables the model to capture the dependencies between feature channels from both average pooling and max pooling perspectives, while suppressing irrelevant channels. A dynamic weighting mechanism is introduced, allowing features at different levels to be adaptively weighted according to their importance during the fusion process, flexibly adjusting the degree of dependence on different features.

[0070] Figure 6 This is a flowchart illustrating the implementation of the hybrid weighted decoder according to an embodiment of the present invention;

[0071] like Figure 6 As shown, features G5 and L5 are first input into the hierarchical fusion module, and a coarse prediction map D6 with one channel is generated through a series of dimensionality reduction convolutions. Then, the hybrid weighted decoder receives F... i (Features from the previous fusion module), D i+1 and D i+2 (Features of adjacent layers) are used as input, and F i With D i+1 D i+2 The features are concatenated and then convolved. After reshaping and reversing, the processed features are added to and multiplied with the original concatenated features. The multiplication operation involves the weights generated by the DSE module.

[0072] Figure 7 This is a flowchart illustrating the implementation of Dynamic Compression and Excitation (DSE) according to an embodiment of the present invention.

[0073] like Figure 7 As shown, the module receives input feature D iThe input features are processed through two branches: average pooling and max pooling. Each branch then proceeds through a fully connected (FC) layer, a ReLU activation function, another FC layer, and a convolutional (Conv) layer. Finally, weights w1 and w2 are generated using the sigmoid function. These two weights are multiplied by the original input feature D, and the weighted results are summed to output the weighted feature D', which is used in the subsequent calculations in the hybrid weighted decoder. After this series of processing steps, the input features undergo another convolution operation and are finally summed with the features processed by the DSE module to output the final decoded feature D. i .

[0074] Step 6: Perform multi-level supervision on the segmentation map and the final prediction map output by the hybrid weighted decoder, and jointly optimize the model parameters through the loss function;

[0075] In this embodiment, the segmentation map D6 is replaced by the output of the previous stage fusion module, and the output of each subsequent decoder is also an input to the next stage decoder. The camouflaged target segmentation map D is generated by concatenating D6 with the four decoders. i,i＝2,3,4,5 The loss is calculated using the ground truth graph (GT) of the training set, and the selected structured loss function is the weighted binary cross-entropy loss. Weighted intersection and combination loss And tensor loss Lc. To handle multi-scale outputs, during training, all outputs D are processed. k＝2,3,4,5,6 Upsampling is performed to match the resolution of the ground truth segment (GT). The formula for the total loss function L of the camouflaged target detection model is as follows:

[0076]

[0077] Where k represents different network layers; D i This represents the segmentation map of camouflaged targets at different layers; GT represents the ground truth label of the camouflaged target.

[0078] Weighted binary cross-entry loss function It has been widely used in the field of image segmentation tasks, and its definition is as follows:

[0079]

[0080] Among them, w ij P represents the weight value of pixel (i,j). ij and G ij The method-predicted value and the ground truth (GT) represent the pixel (i,j) of the camouflaged object. To optimize the global structure, a weighted intersection-over-union (IoU) loss is introduced. It can calculate complete structural similarity. The calculation formula is as follows:

[0081]

[0082] Figure 8 This is a schematic diagram illustrating a qualitative comparison between six mainstream camouflage target segmentation models and the method of this invention in a comparative experiment of embodiments of the present invention.

[0083] Where GT represents the salience label of the camouflaged target, and Our represents the camouflaged target prediction result of the present invention. By comparison, it can be seen that the present invention can effectively detect camouflaged targets, and the edge contour of the camouflaged targets is relatively clear, which improves the accuracy of camouflaged target detection.

[0084] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition, characterized in that: The method includes: S1. Obtain the camouflaged target image, input the camouflaged target image into the encoder network to extract multi-level features, and generate... Feature maps at different scales; S2, will Layer features are input into the global perception module and the local optimization module respectively for feature extraction at different scales, and global key features and local sensitive features are selected to output global features. and local features ; The global perception module consists of several interconnected global visual mamba blocks, used to acquire the relationships between all pixels from a global perspective. Within each global visual mamba block, a dynamic Top-k feature selection mechanism is introduced to filter the most critical global features for the detection task and output the global features. ; The local optimization module consists of several interconnected local visual mamba blocks, used to enhance the spatial local information in the initial features. Within each local visual mamba block, a dynamic Top-k feature selection mechanism is also introduced to dynamically filter key features from the local features and output the local features. This achieves the effect of filtering out background local features that are irrelevant to the target and improving the model's sensitivity to key regions; S3. Global features at the same level and local features Perform cross tensor decomposition, and generate complementary features through low-rank factor matrix expansion and cross-merging; S4. Combine the complementary global features, local features, and previous layer fused features. The input layer fusion module achieves cross-layer feature interaction through gated convolution and dynamic weighting mechanisms, and outputs fused features. ; S5, will The feature input is a hybrid weighted decoder, which combines the channel attention mechanism of the DSE module for inverse optimization, generates and upsamples prediction maps layer by layer, and finally outputs the camouflage target detection result. The input of the latter two decoders not only includes the fused features of the current level, but also the feature maps output by the previous two decoders. S6. Perform multi-level supervision on the segmentation map and the final prediction map output by the hybrid weighted decoder, and jointly optimize the model parameters through the loss function.

2. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, The encoder network in S1 includes: Five initial features were extracted using the ResNet-50 network, where the first layer features... The resolution is greater than the resolution threshold and there is background noise, therefore... Discard, use Complete the subsequent COD missions.

3. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, The global perception module and local optimization module in S2 include: The local optimization module and the global perception module respectively correspond to local Top-k feature selection and global Top-k feature selection, which select the features with the highest relevance to the query. The system uses a single marker to fully utilize sparsity, thereby capturing and extracting the most critical camouflaged target information. in, Dynamically set to a series of values, satisfying ; Indicates the current level index. This represents the total number of levels, and the different levels... The value difference rate is not less than 15%.

4. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 3, characterized in that, The local visual mamba block and the global visual mamba block include: The local visual mamba block adopts a local Top-k feature selection module as a feature extractor. The local visual mamba block contains an unfolding operator, which effectively ensures the spatial proximity of adjacent markers in two-dimensional or three-dimensional arrays. The global visual mamba block uses a global Top-k feature selection module as a feature extractor. After passing through a deep convolutional layer (DWC), an activation function (SiLU), and flattening, it is combined with the Top-k feature selection mechanism to generate features with GRF, which are then forwarded to the next SSM module.

5. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, S3 includes: combining features at the same level and By decomposing the convolutional layers, local information is selectively incorporated into global information to form complementary global features, and vice versa. global features and local features Construct low-rank factor matrices respectively and ; Then, the dimensions of the low-rank factor matrices are respectively expanded to... and ; in, S and T represent the number of input channels and the number of kernels learned by the convolutional layer, respectively; D1 and D2 represent the width and height of the convolutional kernel space window, respectively; and r represents the rank of the original weight matrix before decomposition. Finally, cross-merge to generate the enhanced weight matrix. and Among them, parameter increment .

6. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 5, characterized in that, In order to clearly capture and Complementary cues between layers maximize the inter-layer feature maps output from each branch of the layer. or Distance, and as an additional item in the task training objective; Then, according to the loss formula Lc: ; The distance Lc between the output feature maps is obtained; where p={1,2}, * denotes convolution, x denotes a vector, K, k and ∆K, ∆k have the same dimensions, and represents the convolution filter.

7. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, The hierarchical fusion module in S4 includes: Global weight generation path: GlobalAvgPool → 1x1 convolution → BN → ReLU → 1x1 convolution → Sigmoid; Local weight generation path: dual 1×1 convolutional layer cascaded structure, batch normalization (BN), and GELU loss function; The path for generating pre-stage feature weights is: a concatenated structure of two 1×1 convolutional layers, along with batch normalization and the ReLU loss function; The outputs of the three paths are concatenated, then gated convolution is used to filter redundant information, and residual connections are performed to achieve cross-level feature interaction and output fused features. .

8. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, The hybrid weighted decoder in S5 includes: The hybrid weighted decoder is used to integrate different information by utilizing cross-layer aggregation, dynamic weighting, and inverse optimization: Specifically, the previously fused features and the output segmentation features from adjacent decoders are concatenated, convolved, reshaped, and inverted. The resulting feature map, obtained through addition and multiplication, is then dynamically weighted by the DSE module along with the previous segmentation feature map. Finally, the weighted features are added together to output the segmentation map. ; The DSE module in the hybrid weighted decoder includes a dual attention branch and a weight fusion part. Specifically, the feature map is processed through two branches: average pooling and max pooling. Each branch is then processed through a fully connected layer, a ReLU activation function, another fully connected layer, and a convolutional layer. Finally, the weights are generated using the Sigmoid function.

9. The camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition as described in claim 1, characterized in that, The S6 includes: Segmentation image The output of the previous stage fusion module is used instead, and the output of each subsequent decoder also serves as an input to the next stage decoder. The camouflaged target segmentation map generated by adjacent connections with the four decoders The loss is calculated using the ground truth graph (GT) of the training set, and the selected structured loss function is the weighted binary cross-entropy loss. Weighted average loss and tensor loss ; To handle multi-scale outputs, all outputs are processed during training. Upsampling is performed to match the resolution of the segmented ground truth map GT; The total loss function of the camouflaged target detection model The specific formula is as follows: ； in, Indicates different network layers; This represents a segmentation map of camouflaged targets at different layers; This represents the truth label of the masquerading target.

Images