A method and system for detecting a camouflaged target

Abstract

This invention discloses a method and system for camouflaged target detection, belonging to the field of camouflaged target detection technology. The constructed target detection model outputs multi-scale features of the camouflaged target image through an encoder; a reverse residual preservation module dynamically fuses the multi-scale features of the current batch with the cached multi-scale features of the previous batch to generate enhanced multi-scale features for the current batch; each guided attention dynamic fusion module is used for the extraction and fusion of global semantic and spatial features, outputting the fused features of the current batch at each level of the cascaded architecture; the decoder maps the fused features of the shallowest layer in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result. The dynamic filter kernel is generated by adaptively adjusting the convolution kernel size based on the foreground proportion in the coarsely estimated map obtained from the enhanced multi-scale feature processing. This method can achieve adaptive multi-scale feature fusion, significantly improving localization and boundary accuracy.

Description

Technical Field

[0001] This invention relates to the field of camouflage target detection technology, and more specifically to a camouflage target detection method and system. Background Technology

[0002] Camouflaged object detection (COD) is a relatively new research area that combines image segmentation and salient object detection. Its aim is to accurately detect and segment difficult-to-distinguish camouflaged targets from complex backgrounds. This technology has broad application prospects in fields such as medical image analysis, industrial defect detection, and military reconnaissance. However, the high visual similarity between camouflaged targets and their backgrounds, the diversity of target scale and appearance, and the complexity of target boundary contours greatly increase the difficulty of detection and segmentation, making COD a highly challenging task. This also means that many traditional object detection and segmentation methods are insufficient for COD tasks.

[0003] Traditional methods for detecting camouflaged targets are mostly based on convolutional neural networks, which improve detection performance by combining multi-scale features and boundary information. Subsequent work introduced the Transformer architecture, which uses global context information to model the global semantic information of the hidden target. Although these algorithms have improved COD performance to some extent by enhancing the backbone network, introducing attention mechanisms and adding multi-scale representations.

[0004] However, due to the high visual similarity between camouflaged targets and their backgrounds, the diversity of target scales and appearances, and the complexity of target boundary contours, existing technologies have failed to fully consider the differences in spatial and semantic information in features at various scales, and lack the ability to adaptively fuse spatial and semantic cues, resulting in inaccurate target localization and blurred boundaries. Summary of the Invention

[0005] To address the problems existing in the above-mentioned fields, this invention proposes a camouflaged target detection method and system. The constructed target detection model can achieve adaptive multi-scale feature fusion, significantly improving localization and boundary accuracy.

[0006] To address the aforementioned technical problems, this invention discloses a method and system for detecting camouflaged targets, comprising the following steps: Acquire the image of the disguised target to be detected; The camouflaged target image is input into a pre-trained target detection model, and the camouflaged target detection result is output by prediction. The target detection model uses Transformer as its backbone network, and embeds an inverse residual preservation module and a cascaded self-attention dynamic fusion module sequentially between the encoder and decoder of the original Transformer. The cascaded self-attention dynamic fusion module includes multiple cascaded guided attention dynamic fusion modules, adopting a top-down cascaded architecture. Each guided attention dynamic fusion module includes a feature splicing layer, a semantic branch, a spatial branch, and a dynamic fusion layer. The encoder sequentially outputs the multi-scale features of the current batch of camouflaged target images; the inverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate the enhanced multi-scale features of the current batch; each of the guided attention dynamic fusion modules is used for the extraction and fusion of global semantic features and spatial features, and outputs the fused features of each level in the cascaded architecture; the decoder maps the fused features of each level in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result; The feature concatenation layer concatenates the fused features output from the previous layer and the enhanced multi-scale features of the current layer as the input features of the current layer. The semantic branch extracts the global semantic features of the input features. The spatial branch extracts the spatial features of the input features and multiplies them element-wise with the dynamic filter kernel to obtain refined spatial features. The dynamic fusion layer fuses the global semantic features and refined spatial features to generate the fused features of the current layer. The dynamic filter kernel is generated by adaptively adjusting the size of the convolution kernel based on the foreground proportion in the rough estimate map obtained from the enhanced multi-scale feature processing.

[0007] Preferably, the reverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate enhanced multi-scale features for the current batch, specifically including: The reverse residual preservation module includes a caching and fusion mechanism as well as a dynamic weight graph generator. The caching and fusion mechanism is used to fuse the multi-scale features of the previous batch of cached features with the multi-scale features of the current batch. The dynamic weight graph generator is used to generate a dynamic weight graph, which guides the adaptive balance between the multi-scale features cached in the previous batch and the multi-scale features in the current batch during the fusion process, thereby obtaining the enhanced multi-scale features of the current batch.

[0008] Preferably, the dynamic weight map is obtained by concatenating the multi-scale features of the previous batch and the multi-scale features of the current batch in the channel dimension, and then processing the concatenation result sequentially through the ReLU activation function, convolution operation and Sigmoid activation function.

[0009] Preferably, the feature concatenation layer concatenates the fused features output from the previous layer and the enhanced multi-scale features of the current layer as the input features of the current layer, specifically including: In the feature stitching layer, the fused features output from the previous layer are upsampled to the spatial resolution of the enhanced multi-scale features of the current layer; The fused features from the previous level after unifying the spatial resolution and the enhanced multi-scale features from the current level are concatenated, and the channel dimension is reduced by convolution operation to obtain coarse fused features, which are used as the input features for the current batch.

[0010] Preferably, the semantic branch extracts global semantic features from the input features, specifically including: The semantic branch generates three feature matrices from the input features through a linear mapping. Based on the three feature matrices, global contextual information of the input features is extracted through a self-attention mechanism, and then fused with the input features through residual connections to obtain global semantic features.

[0011] Preferably, the dynamic fusion layer fuses global semantic features and refined spatial features to generate fused features for the current layer, specifically including: The dynamic fusion layer obtains dynamic fusion weights by concatenating global semantic features and refined spatial features and passing them through a Sigmoid activation function. The global semantic features and refined spatial features are weighted and fused using dynamic fusion weights to generate the fused features of the current level.

[0012] Preferably, the target detection model further includes a spatial attention decoder and a learnable filter generator connected in parallel with the cascaded self-attention dynamic fusion module; The spatial attention decoder processes the enhanced multi-scale features of the current batch to generate a single-channel coarse estimation map; The learnable filter generator adaptively adjusts the convolution kernel size based on a rough estimate of the foreground proportion in the image to generate a dynamic filter kernel.

[0013] Preferably, the decoder maps the fused features of each layer in the cascaded architecture into a multi-scale prediction map, which serves as the camouflage target detection result, specifically including: The decoder is a guided refinement decoder, which includes an upsampling module and a partial convolution module in sequence; The upsampling module progressively upsamples the fused features at each level, and the partial convolution module progressively refines the contextual information of the upsampled fused features to generate a multi-scale prediction map. The generated multi-scale prediction map is used as the result of camouflaged target detection.

[0014] Preferably, during the training process, the target detection model employs a deep supervision strategy, which uses a hybrid loss function constructed from a coarse estimation map, a multi-scale prediction map, and a real target image for joint supervision. The hybrid loss function is composed of a weighted combination of binary cross-entropy loss and cross-union ratio loss.

[0015] Preferably, it also includes a camouflage target detection system, comprising: The data acquisition module is used to acquire images of the disguised target to be detected; The target detection module is used to input camouflaged target images into a pre-trained target detection model and output camouflaged target detection results through prediction. The target detection model uses Transformer as its backbone network, and embeds an inverse residual preservation module and a cascaded self-attention dynamic fusion module sequentially between the encoder and decoder of the original Transformer. The cascaded self-attention dynamic fusion module includes multiple cascaded guided attention dynamic fusion modules, adopting a top-down cascaded architecture. Each guided attention dynamic fusion module includes a feature splicing layer, a semantic branch, a spatial branch, and a dynamic fusion layer. The encoder sequentially outputs the multi-scale features of the current batch of camouflaged target images; the inverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate the enhanced multi-scale features of the current batch; each of the guided attention dynamic fusion modules is used for the extraction and fusion of global semantic features and spatial features, and outputs the fused features of each level in the cascaded architecture; the decoder maps the fused features of each level in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result; The feature concatenation layer concatenates the fused features output from the previous layer and the enhanced multi-scale features of the current layer as the input features of the current layer. The semantic branch extracts the global semantic features of the input features. The spatial branch extracts the spatial features of the input features and multiplies them element-wise with the dynamic filter kernel to obtain refined spatial features. The dynamic fusion layer fuses the global semantic features and refined spatial features to generate the fused features of the current layer. The dynamic filter kernel is generated by adaptively adjusting the size of the convolution kernel based on the foreground proportion in the rough estimate map obtained from the enhanced multi-scale feature processing.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The camouflage target detection method proposed in this invention utilizes the enhanced multi-scale features generated by the inverse residual preservation module in the current batch. This effectively leverages the potential correlations across batches, mitigating performance degradation caused by differences in training data distribution, and significantly enhancing the model's training stability and feature consistency. This provides precise feature input for the cascaded self-attention dynamic fusion module. Within the guided attention dynamic fusion module, a feature splicing layer concatenates the fused features from the previous and current layers. Through the collaboration of semantic branches, spatial branches, and the dynamic fusion layer, adaptive fusion of the spatial and semantic information ratios in features at different scales is achieved. This effectively overcomes the problem of insufficient spatial and semantic information fusion in existing technologies, alleviates the challenges posed by changes in target scale, and significantly improves the accuracy of target localization and the segmentation precision of fuzzy boundaries. To further enhance the model's accuracy in predicting and precisely locating the contours of camouflage targets, a dynamically generated dynamic filter kernel provides dynamic positional priors for the spatial features of the spatial branch, guiding the precise refinement of spatial features. This enables the model to possess a strong ability to capture and locate local details of camouflage targets. Attached Figure Description

[0017] Figure 1 This is a flowchart of the camouflage target detection method proposed in this invention; Figure 2 This is the target detection model network architecture provided in the embodiments of the present invention; Figure 3 This is a network structure diagram of the Reverse Residual Preservation Module (RRP) provided in an embodiment of the present invention. Figure 4 This is a network structure diagram of the Guided Attention Dynamic Fusion Module (GADF) provided in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the segmentation results of different models provided in the embodiments of the present invention on various public datasets. Detailed Implementation

[0018] The following will refer to the appendices in the embodiments of the present invention. Figures 1-5 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0019] Example like Figure 1 As shown, this invention proposes a method for detecting camouflaged targets, which includes the following steps: S1: Obtain multiple camouflaged target images, and denote the entire dataset as D, with image size as... .

[0020] S2: Input the acquired camouflaged target image into the pre-trained target detection model, and output the camouflaged target detection result through prediction.

[0021] Among them, the constructed target detection model is as follows Figure 2 As shown, the Transformer is used as the backbone network. The Inverse Residual Preservation Module (RRP) and the Cascaded Self-Attention Dynamic Fusion Module (CSDF) are embedded sequentially between the encoder and decoder of the original Transformer. The CSDF includes multiple cascaded Guided Attention Dynamic Fusion Modules (GADF), which adopt a top-down cascaded architecture. Each GADF includes a feature concatenation layer, a semantic branch, a spatial branch, and a dynamic fusion layer.

[0022] S21: Input the camouflaged target image into the backbone network. The backbone network gradually constructs the current batch of images through multi-head self-attention (MSA) and Mix modules. j Multiscale features ,in This indicates different scale level indexes.

[0023] like Figure 2 As shown in the figure, this is the network architecture of the target detection model constructed in this invention. -- p1, p2, p3, and p4 represent the fused features corresponding to scales 1 to 4, respectively. Figure 2 The English text in the bottom right corner, from top to bottom, reads: element-wise addition operation, spatial attention decoder, inverse residual preservation, guided attention dynamic fusion module, and deep supervision.

[0024] like Figure 3 The diagram shown is a network structure diagram of the RRP module. Figure 3 In this context, "batch" refers to the batch of samples input to the network. The RRP module is integrated into each training iteration. Its core idea is to use a reverse residual strategy, where the RRP module extracts the previous batch of camouflaged target images from the encoder. Enhanced multi-scale features Multi-scale features of the current batch Perform feature fusion to obtain the first Scale in the current batch Enhanced features for:

[0025] ; in, This represents a dynamic weighting graph, used to guide the adaptive balance between historical and current features. express and The integration This represents a convolution-batch normalized-ReLU sequence. For a full 1 tensor, This represents element-wise multiplication; Dynamic weight graph for: ; in, It is the Sigmoid activation function. yes Convolution operation, and These are the ReLU activation function and the concatenation of features along the channel dimension, respectively.

[0026] This process can maintain the intrinsic relationships between similar samples in the training set and make full use of the potential correlations between batches, thereby enhancing the stability and consistency of features.

[0027] S22: To better incorporate prior knowledge of the camouflaged target's location, the target detection model constructed in this invention also includes a spatial attention decoder (SAD) and a filter generation module (LFG) that run parallel to the CSDF module, wherein: SAD enhances multi-scale features After resampling and stitching, through Convolutions and sigmoid activations are used to generate a spatial attention map, which is then decoded into a coarse estimate map for a single channel. ; LFG module according to Prospect ratio in Adaptively adjust the convolution kernel size to generate dynamic filter kernels. : ; in, It is a round-down operation.

[0028] Will As a location prior, it is used to guide the feature refinement and localization of the feature extraction process in spatial branching.

[0029] S23: Enhance the multi-scale features of the current batch output by the RRP module. Pass it to CSDF.

[0030] The CSDF module adopts a top-down cascaded architecture, consisting of four Guided Attention Dynamic Fusion (GADF) modules, which enable the orderly propagation of high-level semantic information to low-level features. Each GADF module includes a feature concatenation layer, a semantic branch, a spatial branch, and a dynamic fusion layer.

[0031] like Figure 4 As shown, in the GADF module, the feature concatenation layer combines the fused features output from the previous layer. Multi-scale features enhanced at the current level Concatenate the features as input for the current layer. Specifically: The feature splicing layer first fuses the features Upsampling to Spatial resolution; and the fused features after unifying resolution and To splice together, and through a Convolution reduces channel dimensions, thus obtaining the input features of the current layer. for: ; Then, the input features of the current level are... Simultaneously, the data is fed into the semantic branch and the spatial branch to extract the semantic features and spatial features of the input features at the current level, respectively, where: In the semantic branch, the input features of the current level Mapped to and Then, a self-attention mechanism is employed, and residual connections are used to... Fusion to obtain global semantic features for: ; In the spatial branch, the dynamic filter kernel is used as a positional prior to guide feature extraction in the spatial branch; for the input features of the current level After channel dimensionality reduction, the mapping is as follows: and Then, a self-attention mechanism is employed, and residual connections are used to... The features are fused and multiplied element-wise with the dynamic filter kernel to obtain the spatial features of the input features at the current level. for: ; in, This represents the dynamic filter kernel output by the filter generation module; sft (.) represents the softmax operation. rsp (.) represents the reshape operation.

[0032] In dynamic fusion, the dynamic fusion layer fuses global semantic features and refined spatial features to generate fused features at the current scale. Dynamic fusion weights Integration features with the current level They are respectively: ; ; Fusion features for: ; in, Indicates the first Fusion features of outputs at each level , This represents the multi-scale features enhanced at the current level. This represents a rough estimate. This represents the fused features of the output from the previous level.

[0033] The current level of fusion features , as input features for the next level.

[0034] This mechanism can effectively integrate shallow texture details with deep semantic information, alleviate feature differences, and enhance the model's adaptability to camouflaged targets at various scales.

[0035] S24: The fused features of each level in the cascaded architecture are passed to the decoder. The decoder maps the fused features of each level into a multi-scale prediction map, which serves as the result of camouflaged target detection.

[0036] In this embodiment, the decoder is a Guided Refinement Decoder (GRD). GRD utilizes high-level semantic features to progressively guide the fusion of low-level detailed features, and embeds a partial Convolutional Processing Module (PCM) during the decoding process to refine contextual information and generate a multi-scale prediction map. p i , i =1,2,3,4; and the multi-scale prediction map p i This serves as the result of camouflaged target detection.

[0037] S25: Model Training This invention employs a deep supervision strategy, wherein, based on a coarse estimation graph... Multi-scale prediction map We construct a hybrid loss function using real target images and perform joint supervision.

[0038] The hybrid loss function consists of binary cross-entropy loss ( ) and crossover ratio loss ( Together, they form a system that enables coordinated optimization of the model at the pixel level and the foreground-background region level, thereby improving sensitivity to complex contours and overall segmentation accuracy.

[0039] The hybrid loss function constructed during training is as follows: ; in, G The label represents the Ground Truth, i.e., the real target image. These are learnable weight parameters.

[0040] Multiple sets of experiments have verified that the method proposed in this invention has significant advantages over traditional camouflaged target detection methods: In the target detection model constructed in this invention, the RRP module effectively utilizes the potential correlation across batches, alleviating the performance degradation caused by differences in training data distribution, and greatly enhancing the training stability and feature consistency of the model. In the Cascaded Self-Attention Dynamic Fusion (CSDF) architecture, the Guided Attention Dynamic Fusion (GADF) module achieves adaptive fusion of spatial and semantic information ratios. Combined with the dynamic filter kernel generated by the learnable spatial attention decoder and filter generator (LFG), it provides dynamic positional priors for spatial feature extraction in the Guided Attention Dynamic Fusion module, making the target detection model more accurate in predicting the contours and precisely locating camouflaged targets. By balancing the two through dynamic adaptive fusion, the model can grasp the overall semantics while also depicting fine boundaries, achieving an orderly and progressive fusion of high-level semantics to low-level details. Information flow is more efficient, enhancing the model's ability to interact with multi-scale features and adapt to targets of different scales.

[0041] Ultimately, in complex camouflaged target detection tasks, the method of this invention outperforms most existing camouflaged target detection methods in terms of target localization precision and model robustness, verifying its great potential in practical application scenarios.

[0042] Figure 5 This diagram illustrates the segmentation results of the proposed method and other comparative models on the COD10K, NC4K, and CAMO datasets. It sequentially displays the original image, ground truth annotations, and detection results from different methods. Compared to other methods, the segmentation mask generated by the proposed method closely matches the boundaries of the real target, effectively handling blurred and fragmented contours and reducing over-segmentation or under-segmentation. Therefore, this invention achieves superior performance in delineating the boundaries and locating camouflaged targets, thereby improving the overall detection accuracy of the model.

[0043] This invention also proposes a camouflaged target detection system, comprising: The data acquisition module is used to acquire images of the disguised target to be detected; The target detection module is used to input camouflaged target images into a pre-trained target detection model and output camouflaged target detection results through prediction. The object detection model uses Transformer as the backbone network, and embeds an inverse residual preservation module and a cascaded self-attention dynamic fusion module sequentially between the encoder and decoder of the original Transformer. The cascaded self-attention dynamic fusion module includes multiple cascaded guided attention dynamic fusion modules, adopting a top-down cascaded architecture. Each guided attention dynamic fusion module includes a feature concatenation layer, a semantic branch, a spatial branch, and a dynamic fusion layer. The encoder sequentially outputs the multi-scale features of the current batch of camouflaged target images; the inverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate the enhanced multi-scale features of the current batch; each guided attention dynamic fusion module is used for the extraction and fusion of global semantic features and spatial features, and outputs the fused features of each level in the cascaded architecture; the decoder maps the fused features of each level in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result; The feature concatenation layer concatenates the fused features output from the previous layer with the enhanced multi-scale features of the current layer as the input features of the current layer; the semantic branch extracts the global semantic features of the input features; the spatial branch extracts the spatial features of the input features and multiplies them element-wise with the dynamic filter kernel to obtain refined spatial features; the dynamic fusion layer fuses the global semantic features and refined spatial features to generate the fused features of the current layer; the dynamic filter kernel is generated by adaptively adjusting the size of the convolution kernel based on the foreground proportion in the rough estimate map obtained from the enhanced multi-scale feature processing.

[0044] In the target detection model constructed by the method proposed in this invention, the embedded RRP module preserves the continuity of data distribution by fusing multi-scale features from the previous batch, making model learning smoother, alleviating the feature distribution differences caused by batch partitioning, and improving training efficiency and generalization ability. The GADF module enables orderly and progressive fusion of high-level semantics to low-level details, making information flow more efficient and enhancing model robustness and multi-scale adaptability. In the GADF module, the global context is captured through semantic branches to understand "what is the target"; local details are focused through spatial branches to understand "where the target is". The dynamic filter kernel generated by the spatial attention decoder and filter generation module, which are parallel to the CSDF module, is used as the dynamic position prior of spatial features in the GADF module. Through dynamic fusion, the global context and local details are adaptively balanced to achieve adaptive multi-scale feature fusion. This allows the model to grasp both the overall semantics and the fine boundaries, significantly improving localization and boundary accuracy, especially adapting to the challenges of camouflaged targets with varying scales and appearances.

[0045] The method proposed in this invention generates a coarse estimation map through SAD, providing prior knowledge of "where the target may be". The LFG module dynamically generates filter kernels based on the prior knowledge, which strengthens the features of key regions of the target in the spatial branch, greatly enhancing the model's ability to perceive local details such as the target center and contour, and making the segmentation results more compact and accurate.

[0046] During the training phase, deep supervision is used to jointly optimize the multi-scale prediction graph, ensuring that each layer of the network produces effective output, accelerating convergence and improving the accuracy of the final fusion result.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0048] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for detecting camouflaged targets, characterized in that, Includes the following steps: Acquire the image of the disguised target to be detected; The camouflaged target image is input into a pre-trained target detection model, and the camouflaged target detection result is output by prediction. The target detection model uses Transformer as its backbone network, and embeds an inverse residual preservation module and a cascaded self-attention dynamic fusion module sequentially between the encoder and decoder of the original Transformer. The cascaded self-attention dynamic fusion module includes multiple cascaded guided attention dynamic fusion modules, adopting a top-down cascaded architecture. Each guided attention dynamic fusion module includes a feature splicing layer, a semantic branch, a spatial branch, and a dynamic fusion layer. The encoder sequentially outputs the multi-scale features of the current batch of camouflaged target images; the inverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate the enhanced multi-scale features of the current batch; each of the guided attention dynamic fusion modules is used for the extraction and fusion of global semantic features and spatial features, and outputs the fused features of each level in the cascaded architecture; the decoder maps the fused features of each level in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result; The feature concatenation layer concatenates the fused features output from the previous layer and the enhanced multi-scale features of the current layer as the input features of the current layer. The semantic branch extracts the global semantic features of the input features. The spatial branch extracts the spatial features of the input features and multiplies them element-wise with the dynamic filter kernel to obtain refined spatial features. The dynamic fusion layer fuses the global semantic features and refined spatial features to generate the fused features of the current layer. The dynamic filter kernel is generated by adaptively adjusting the size of the convolution kernel based on the foreground proportion in the rough estimate map obtained from the enhanced multi-scale feature processing.

2. The camouflaged target detection method according to claim 1, characterized in that, The reverse residual preservation module dynamically fuses the multi-scale features of the current batch with the cached multi-scale features of the previous batch to generate enhanced multi-scale features for the current batch, specifically including: The reverse residual preservation module includes a caching and fusion mechanism as well as a dynamic weight graph generator. The caching and fusion mechanism is used to fuse the multi-scale features of the previous batch of cached features with the multi-scale features of the current batch. The dynamic weight graph generator is used to generate a dynamic weight graph, which guides the adaptive balance between the multi-scale features cached in the previous batch and the multi-scale features in the current batch during the fusion process, thereby obtaining the enhanced multi-scale features of the current batch.

3. The camouflaged target detection method according to claim 2, characterized in that, The dynamic weight map is obtained by concatenating the multi-scale features of the previous batch and the multi-scale features of the current batch in the channel dimension, and then processing the concatenation result through the ReLU activation function, convolution operation and Sigmoid activation function in sequence.

4. The camouflaged target detection method according to claim 1, characterized in that, The feature concatenation layer concatenates the fused features output from the previous layer with the enhanced multi-scale features of the current layer, using this as the input features for the current layer. Specifically, it includes: In the feature stitching layer, the fused features output from the previous layer are upsampled to the spatial resolution of the enhanced multi-scale features of the current layer; The fused features from the previous level after unifying the spatial resolution and the enhanced multi-scale features from the current level are concatenated, and the channel dimension is reduced by convolution operation to obtain coarse fused features, which are used as the input features for the current batch.

5. The camouflaged target detection method according to claim 1, characterized in that, The semantic branch extracts global semantic features from the input features, specifically including: The semantic branch generates three feature matrices from the input features through a linear mapping. Based on the three feature matrices, global contextual information of the input features is extracted through a self-attention mechanism, and then fused with the input features through residual connections to obtain global semantic features.

6. The camouflaged target detection method according to claim 1, characterized in that, The dynamic fusion layer fuses global semantic features and refined spatial features to generate fused features for the current layer, specifically including: The dynamic fusion layer obtains dynamic fusion weights by concatenating global semantic features and refined spatial features and passing them through a Sigmoid activation function. The global semantic features and refined spatial features are weighted and fused using dynamic fusion weights to generate the fused features of the current level.

7. The camouflaged target detection method according to claim 1, characterized in that, The target detection model also includes a spatial attention decoder and a learnable filter generator connected in parallel with the cascaded self-attention dynamic fusion module; The spatial attention decoder processes the enhanced multi-scale features of the current batch to generate a single-channel coarse estimation map; The learnable filter generator adaptively adjusts the convolution kernel size based on a rough estimate of the foreground proportion in the image to generate a dynamic filter kernel.

8. The camouflaged target detection method according to claim 1, characterized in that, The decoder maps the fused features from each layer of the cascaded architecture into a multi-scale prediction map, which serves as the result of camouflaged target detection. Specifically, this includes: The decoder is a guided refinement decoder, which includes an upsampling module and a partial convolution module in sequence; The upsampling module progressively upsamples the fused features at each level, and the partial convolution module progressively refines the contextual information of the upsampled fused features to generate a multi-scale prediction map. The generated multi-scale prediction map is used as the result of camouflaged target detection.

9. The camouflaged target detection method according to claim 1, characterized in that, During the training process, the target detection model adopts a deep supervision strategy, which uses a hybrid loss function constructed from a coarse estimation map, a multi-scale prediction map, and a real target image for joint supervision. The hybrid loss function is composed of a weighted combination of binary cross-entropy loss and cross-union ratio loss.

10. A camouflaged target detection system, characterized in that, include: The data acquisition module is used to acquire images of the disguised target to be detected; The target detection module is used to input camouflaged target images into a pre-trained target detection model and output camouflaged target detection results through prediction. The target detection model uses Transformer as its backbone network, and embeds an inverse residual preservation module and a cascaded self-attention dynamic fusion module sequentially between the encoder and decoder of the original Transformer. The cascaded self-attention dynamic fusion module includes multiple cascaded guided attention dynamic fusion modules, adopting a top-down cascaded architecture. Each guided attention dynamic fusion module includes a feature splicing layer, a semantic branch, a spatial branch, and a dynamic fusion layer. The encoder sequentially outputs the multi-scale features of the current batch of camouflaged target images; the inverse residual preservation module dynamically fuses the multi-scale features of the current batch with the multi-scale features cached in the previous batch to generate the enhanced multi-scale features of the current batch; each of the guided attention dynamic fusion modules is used for the extraction and fusion of global semantic features and spatial features, and outputs the fused features of each level in the cascaded architecture; the decoder maps the fused features of each level in the cascaded architecture to a multi-scale prediction map as the camouflaged target detection result; The feature concatenation layer concatenates the fused features output from the previous layer and the enhanced multi-scale features of the current layer as the input features of the current layer. The semantic branch extracts the global semantic features of the input features. The spatial branch extracts the spatial features of the input features and multiplies them element-wise with the dynamic filter kernel to obtain refined spatial features. The dynamic fusion layer fuses the global semantic features and refined spatial features to generate the fused features of the current layer. The dynamic filter kernel is generated by adaptively adjusting the size of the convolution kernel based on the foreground proportion in the rough estimate map obtained from the enhanced multi-scale feature processing.

Images