An unsupervised camouflage target detection method based on self-anchored progressive learning

Abstract

This invention provides an unsupervised camouflage target detection method based on self-anchored progressive learning, comprising: Step S1, acquiring the image to be processed and inputting it into a self-supervised feature extraction network to extract multi-layer semantic features; Step S2, constructing a semantically driven region detector based on the multi-layer semantic features and enhancing the foreground and background separability to generate a region activation map; Step S3, performing multi-scale fusion on the region activation map and using dual-threshold filtering to obtain foreground anchor regions and background anchor regions, and setting uncertain regions between thresholds as ignored regions to form region-level anchor labels; Step S4, constructing a progressive contextual reasoning network based on the region-level anchor labels, fusing multi-layer semantic features and performing cross-scale interactive reasoning to generate camouflage target detection results. The beneficial effect is that this invention can achieve unsupervised camouflage target detection without manual annotation, and while suppressing the propagation of false label noise, it also ensures detection stability and accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of image target detection, and more specifically, to an unsupervised camouflaged target detection method based on self-anchored progressive learning. Background Technology

[0002] Camouflage target detection aims to identify targets in complex natural scenes that are highly similar to the background in color, texture, and structure. It is widely used in scenarios such as ecological monitoring, military reconnaissance, and security surveillance. Compared with regular salient targets, camouflage targets usually exhibit characteristics such as weak boundaries, low contrast, and local structural breaks, making detection models more reliant on high-level semantics and contextual reasoning capabilities, significantly increasing the task difficulty.

[0003] In recent years, deep learning methods have driven improvements in the performance of camouflaged object detection (COD). However, most methods rely on pixel-level labels for supervised training, resulting in high costs for data acquisition and fine annotation, which hinders scalability. To reduce annotation costs, current research has begun to explore unsupervised methods for camouflaged object detection.

[0004] However, in camouflaged scenarios, due to the high similarity between the target and the background and the blurred boundaries, existing unsupervised camouflaged target detection methods often generate pseudo-labels with problems such as incomplete foreground regions, false activation of the background, and boundary offsets, leading to a significant increase in pseudo-supervision noise. Furthermore, the training process based on pixel-level pseudo-labels easily allows noise to propagate and accumulate, causing instability in the detection model during training and defects such as fragmented detection results and inaccurate boundaries. In addition, some methods introduce complex multi-stage iterations or additional prior modules to mitigate noise, increasing the complexity of training and inference, and still failing to achieve both stability and accuracy under unlabeled conditions. Therefore, there is an urgent need for an unsupervised camouflaged target detection method that can reduce the propagation of pseudo-label noise, improve the integrity and boundary accuracy of the target structure, and enhance the stability of model training and detection without manual annotation, in order to overcome the above-mentioned shortcomings. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to achieve unsupervised camouflage target detection without manual annotation, and to ensure detection stability and accuracy while suppressing the propagation of false label noise. To overcome the shortcomings of the above-mentioned prior art (or related technologies), this invention provides an unsupervised camouflage target detection method based on self-anchored progressive learning.

[0006] This invention provides an unsupervised camouflaged target detection method based on self-anchored progressive learning, comprising the following steps: Step S1: Obtain at least one image to be processed and input it into a pre-constructed self-supervised feature extraction network to extract multi-layer semantic features; Step S2: Based on the multi-layer semantic features, a semantically driven region detector is constructed, and cross-image contrastive learning is used to enhance the foreground and background separability to generate a region activation map. Step S3: Perform multi-scale fusion on the region activation map and use dual threshold filtering to obtain the foreground anchor region and background anchor region, and set the uncertain region between the thresholds as the ignored region to form region-level anchor label; Step S4: Based on the region-level anchor labels, a progressive contextual reasoning network is constructed, which integrates the multi-layer semantic features and performs cross-scale interactive reasoning. Under anchor region supervision, the structure of the camouflaged target is expanded and refined based on the foreground anchor region and the background anchor region to generate camouflaged target detection results.

[0007] The unsupervised camouflaged target detection method based on self-anchored progressive learning proposed in this invention has the following advantages compared with existing technologies: In this invention, the image to be processed is acquired and multi-layer semantic features are extracted in step S1. In step S2, a semantically driven region detector is constructed and a region activation map is generated using cross-image contrastive learning. In step S3, the region activation map is fused at multiple scales to generate region-level anchor labels. In step S4, a progressive contextual reasoning network is constructed and cross-scale angle reasoning is used to refine the structure of the camouflaged target and generate the camouflaged target detection result. The entire process can achieve unsupervised camouflaged target detection without the need for manual annotation. Furthermore, the use of a dual-threshold screening method to divide the foreground anchor region and the background anchor region can suppress the propagation of false label noise. By setting the anchor region supervision conditions, the structure of the camouflaged target can be gradually expanded and refined, improving the stability and accuracy of the detection.

[0008] In one possible implementation, step S1, before inputting the image to be processed into the self-supervised feature extraction network, further includes: The image to be processed is subjected to size normalization processing, and the image to be processed is scaled to 320x320 pixels using the LANCZOS resampling method.

[0009] Compared with existing technologies, the above technical solution can unify the size and sampling method of the image to be processed, reduce feature shift caused by different resolutions and interpolation differences, facilitate batch training and stable inference, and improve the consistency of feature extraction.

[0010] In one possible implementation, step S1 uses a Vision Transformer self-supervised pre-trained model as the self-supervised feature extraction network, and the multi-layer semantic features include at least the last three layers of output features of the Vision Transformer self-supervised pre-trained model.

[0011] Compared with existing technologies, the above-mentioned technical solution can utilize the Vision Transformer self-supervised pre-trained model obtained by pre-training on large-scale unlabeled data to produce general semantic representations as multi-layer semantic features, thereby improving the reliability of semantic features under unsupervised conditions.

[0012] In one possible implementation, in step S2, the semantically driven region detector includes a feature transformation unit and parallel left, right, and summation branches. Semantic consistency enhancement processing is performed on the branch features extracted from each of the left, right, and summation branches. The semantic consistency enhancement processing includes: An attention map is generated based on the branch features. An uncertainty modulation map is generated by multiplying the attention map and its complement map element by element. The processed uncertainty modulation map is then fused with the branch features element by element to obtain the enhanced branch features.

[0013] Compared with existing technologies, the above-mentioned technical solution can enhance feature diversity and robustness through multi-branch parallel modeling, making the regional response more stable and reducing the sensitivity of a single path to noise and local camouflage textures; it can also explicitly highlight the discrimination information of uncertain regions and adaptively enhance branch features, improve the representation ability of boundaries and easily confused regions, thereby improving the reliability of the region activation map.

[0014] In one possible implementation, in step S2, the enhanced branch features of each branch are concatenated along the channel dimension and then further convolved and the semantic consistency enhancement processing is performed again to output a region feature map. Subsequently, a 3x3 convolution is performed on the region feature map through a decoupling head to obtain a single-channel activation map. Based on the single-channel activation map and its complement map, the region feature map is weighted and converged to obtain foreground prototype features and background prototype features.

[0015] Compared with existing technologies, the above-mentioned technical solution can obtain purer foreground and background representative features at the regional level, improve the quality of foreground and background separation, and provide stable prototype support for cross-image contrastive learning.

[0016] In one possible implementation, in step S2, positive sample pairs are formed by cross-image contrastive learning using the foreground prototype features corresponding to different images or the background prototype features corresponding to different images, and negative sample pairs are formed using the foreground prototype features and the background prototype features of the same image, in order to enhance the foreground and background separability and generate the region activation map.

[0017] Compared with existing technologies, the above-mentioned technical solution can achieve cross-border... Figure 1 Consistency constraints enhance the foreground and background discrimination boundary, reduce interference from camouflaged backgrounds, and improve the generalization and stability of region activation maps in different scenarios.

[0018] In one possible implementation, in step S3, multi-scale fusion processes the region activation map using multiple preset scales to obtain the activation value of each region. Dual-threshold filtering uses preset foreground and background thresholds. Regions with activation values ​​greater than the foreground threshold are marked as foreground anchor regions, regions with activation values ​​less than the background threshold are marked as background anchor regions, and regions with activation values ​​between the foreground and background thresholds are marked as ignored regions.

[0019] Compared with existing technologies, the above-mentioned technical solution can improve the stability of the region activation map under multiple preset scales, and significantly reduce the propagation of false label noise caused by uncertain regions by dividing the anchor region and the ignored region, thereby improving the reliability of the false supervision signal.

[0020] In one possible implementation, in step S4, the anchor point region supervision condition is that supervision loss is calculated only for the foreground anchor point region and the background anchor point region, and supervision loss is not calculated for the ignored region.

[0021] Compared with existing technologies, the above technical solution can avoid misleading updates caused by using uncertain regions as supervision signals in optimization, thereby improving the stability and convergence reliability of the training process.

[0022] In one possible implementation, the cross-scale interactive reasoning in step S4 includes: The feature channels are divided into multiple groups, and dilated convolutions with different dilation rates are used to aggregate the multi-layer semantic features within and between groups to output fused features. Then, adaptive average pooling is performed on the fused features to generate a global description, and the global description is used to generate attention weights through 1x1 convolution and nonlinear mapping. The attention weights are then used to weight the current semantic features and residual information respectively before fusion.

[0023] Compared with existing technologies, the above-mentioned technical solution can perform contextual interaction and information propagation under different receptive fields, enhance the multi-scale modeling capability of camouflaged target structure and boundary, thereby improving the structure recovery effect; and can introduce global semantic guidance to suppress irrelevant responses and highlight the relevant context of camouflaged targets, thereby improving the stability of the reasoning process and the detection accuracy. Attached Figure Description

[0024] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a schematic diagram of the framework for unsupervised camouflaged target detection of the present invention; Figure 3 This is a schematic diagram of the framework of the semantic-driven region detector of the present invention; Figure 4 This is a schematic diagram illustrating the generation of the regional anchor point label according to the present invention; Figure 5 This is a schematic diagram of the framework of the progressive contextual reasoning network of the present invention; Figure 6 This diagram illustrates a comparison of the detection performance of the method of this invention with existing unsupervised camouflaged target detection methods. Detailed Implementation

[0025] First, those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0027] See Figure 1 and Figure 2 This invention discloses an unsupervised camouflaged target detection method based on self-anchored progressive learning, comprising the following steps: Step S1: Obtain at least one image to be processed and input it into a pre-constructed self-supervised feature extraction network to extract multi-layer semantic features; Step S2: A semantically driven region detector is constructed based on multi-layer semantic features, and a cross-image contrastive learning method is used to enhance the foreground and background separability to generate region activation maps. Step S3: Perform multi-scale fusion on the region activation map and use dual threshold filtering to obtain the foreground anchor region and background anchor region, and set the uncertain region between the thresholds as the ignored region to form region-level anchor label; Step S4: A progressive contextual reasoning network is constructed based on regional anchor labels. Multi-layer semantic features are fused and cross-scale interactive reasoning is performed. Under anchor region supervision, the structure of the camouflaged target is expanded and refined based on the foreground anchor region and the background anchor region to generate camouflaged target detection results.

[0028] In this embodiment of the invention, before inputting the image to be processed into the self-supervised feature extraction network in step S1, the method further includes: The image to be processed is normalized in size. The LANCZOS resampling method is used to scale the image to 320×320 pixels. The LANCZOS resampling method can effectively preserve the edge details and texture information of the image. Compared with bilinear interpolation or nearest neighbor interpolation, its sampling quality is higher. It can reduce the feature distribution offset problem caused by different resolution image inputs, facilitate batch training and stable inference, and improve the consistency of feature extraction.

[0029] In this embodiment of the invention, in step S1, the image to be processed is first acquired. This image can come from multiple image sets of different scenes, containing various types of camouflaged target images. A Vision Transformer self-supervised pre-trained model is used as the self-supervised feature extraction network. In one specific implementation, DINO (DIstillation with NO) is employed. The ViT-S / 8 version, pre-trained under the label framework, serves as the backbone network for feature extraction. Its architecture includes a 12-layer Transformer encoder, with each layer containing 384 hidden features and 3 heads per layer. A self-supervised feature extraction network extracts multi-layer semantic features from the image to be processed. During this process, the input image is segmented by the DINO model, converting the original 320×320 pixel image into a 40×40 pixel feature representation. Each feature map represents the semantic information of a local region in the image. The multi-layer semantic features extracted by this self-supervised feature extraction network not only contain low-level texture information but also capture high-level semantic information. After processing, the semantic information in the image is compressed into a series of feature maps, each containing a local semantic representation of the image. This process utilizes a multi-layer self-attention mechanism to effectively extract and integrate the contextual information of the image. Each layer progressively extracts different levels of semantic features, continuously improving from low-level details to high-level semantics. The final output 40×40 feature map contains rich semantic information, providing reliable support for subsequent camouflage target detection tasks.

[0030] In this embodiment of the invention, in step S2, a multi-layer semantic feature extracted in step S1 is constructed as follows: Figure 3The semantic-driven region detector (SDR) shown first inputs the multi-layer semantic features extracted in step S1 into the SDR. Its main task is to identify high-confidence foreground and background in the image and generate corresponding region activation maps. Specifically, the SDR processes the input feature map through a convolutional neural network (CNN). This CNN consists of multiple convolutional layers and batch normalization layers. Each convolutional layer uses a 3x3 kernel with a stride of 1 and padding of 1. The number of channels after convolution decreases sequentially from 384 to 128. Through multi-layer convolutional operations, high-level semantic information of the image is extracted step by step. In this process, the convolutional layers can capture key regions in the image and perform non-linear transformations through activation functions, thereby highlighting foreground anchor regions and suppressing... Background interference is addressed by enhancing the convolutional feature map through a self-attention mechanism. This enhancement is achieved through three parallel branches. Each branch first uses a 1x1 convolutional layer to increase the number of channels to 128. After each convolutional layer, a batch normalization layer and activation function are applied to generate a new feature map Z. This feature map Z is then processed by the sigmoid function to generate an attention map, which is then inverted and multiplied pixel by pixel. This attention map is then multiplied pixel by pixel with the original feature map Z and added to it. Finally, the output feature maps from the three branches are concatenated according to the channel dimension to generate a 384-channel feature map. The concatenated feature map is then subjected to the self-attention mechanism again to further improve the separability between the foreground anchor region and the background anchor region, thus generating the final feature map.

[0031] In this embodiment of the invention, the final feature map generated above is used to calculate the region activation map. The region activation map is obtained by pooling and normalizing the final feature map to obtain the activation value of each region, which represents the probability that the region belongs to the foreground or background. In order to enhance the separability between the foreground and the background, this embodiment adopts a cross-image contrastive learning strategy. By calculating the feature similarity between different images, the activation value of the foreground anchor region is strengthened, while the activation value of the background anchor region is weakened. Specifically, the feature map is optimized by contrastive loss so that the foreground anchor region is more prominent in all images, while the activation value of the background anchor region is weakened as much as possible. In the first stage of training, the semantic-driven region detector is trained for 40 rounds with a batch size of 32. The optimizer adopts stochastic gradient descent and the learning rate is set to 1×10. -4 The momentum is set to 0.9, and the weight decay is 1×10. -4 In the loss function, the weight coefficients Set the value to 1, and set other loss weights to 0 to ensure the stability of feature extraction as the main optimization objective during training. Through the above operations, the semantically driven region detector can generate a high-quality region activation map, where each activation value corresponds to a region in the image and is used to indicate whether the region contains a camouflaged target. Finally, the region activation map will serve as a supervision signal for the target region, providing reliable support for subsequent camouflaged target detection tasks.

[0032] See Figure 4 In this embodiment of the invention, in step S3, the region activation map generated in step S2 is first subjected to multi-scale fusion processing to enhance the detail information of the foreground anchor region and the background anchor region in the image. Specifically, the region activation map is processed by convolution operations of different preset scales so that the image features are fully captured at multiple preset scales, thereby obtaining richer semantic information and detail representation. The feature map of each preset scale is sized by upsampling or downsampling, which are 0.5, 1, 1.5, and 2 respectively, to ensure that the feature maps of different preset scales can be effectively fused. After weighted fusion, a fused feature map is generated. The fused feature map contains richer spatial information and contextual information, which helps to improve the separability of the foreground anchor region and the background anchor region, especially the ability to identify camouflaged targets in complex backgrounds.

[0033] In this embodiment of the invention, after multi-scale fusion is completed, the fused feature map is subjected to dual-threshold screening to generate foreground anchor regions and background anchor regions. For this purpose, two parameters, foreground threshold and background threshold, are set to measure the probability that each pixel belongs to the foreground or background. The foreground threshold is set to 0.6 and the background threshold is set to 0.2. When the activation value of a pixel is greater than the foreground threshold, the pixel position is marked as a foreground anchor region. When the activation value of a pixel is less than the background threshold, the pixel position is marked as a background anchor region. Uncertain regions with activation values ​​between the foreground threshold and the background threshold are marked as ignored regions because there is a large uncertainty in the determination of foreground and background. These pixels do not participate in the loss calculation in the subsequent training process, thereby avoiding negative impact on the training of the semantically driven region detector.

[0034] In this embodiment of the invention, through the above-mentioned dual-threshold filtering operation, the final generated region-level anchor label includes a foreground anchor region, a background anchor region, and an ignored region. This region-level anchor label serves as a supervision signal during the training process, providing accurate guidance for subsequent camouflage target detection tasks. This enables the semantically driven region detector to more effectively distinguish between the foreground anchor region and the background anchor region, thereby improving the detection accuracy and stability of camouflage targets.

[0035] In this embodiment of the invention, in step S4, a region-level anchor tag generated in step S3 is used to construct a structure as follows: Figure 5The progressive context inference network shown first inputs region-level anchor labels as supervision signals. The main task of this progressive context inference network is to gradually expand and refine the structure of the camouflaged target through the fusion of multi-layer semantic features. In the specific implementation process, the progressive context inference network uses foreground and background anchor regions to aggregate feature maps from different levels. Through convolution operations and residual connections, features at different preset scales are fused to obtain more accurate global context information. This process effectively integrates shallow and deep features, making the boundaries of the camouflaged target clearer, while also suppressing background interference.

[0036] In this embodiment of the invention, during the cross-scale interactive reasoning process, the progressive context reasoning network gradually refines the foreground anchor point region through alternating reasoning modules, and expands the structure of the camouflaged target layer by layer by combining global context information. The reasoning module adopts a cross-scale feature fusion strategy and enhances the information transmission between different preset scales through operations such as dilated convolution. There are four dilated convolutions, with a kernel size of 3x3 and dilation rates of 1, 2, 2, and 4, respectively, so that the camouflaged target can be more accurately located and segmented at various resolutions.

[0037] In this embodiment of the invention, during the contextual inference process, the progressive contextual inference network gradually expands the boundary of the target region based on the input region-level anchor labels, and finally outputs a refined camouflage target detection result. Each layer's inference process refines the boundary of the target region under the guidance of the foreground anchor region, enabling the progressive contextual inference network to gradually recover the camouflage target structure in the image. Through continuous refinement and optimization, it can provide accurate target region location and fine boundary information, ensuring accurate identification of camouflage targets and effectively avoiding detection errors caused by background interference or target blurring. This results in excellent and robust camouflage target detection results. In the second stage of training, the main loss function is a partial binary cross-entropy loss. The calculation formula is as follows: in, This represents the region-level anchor labels from the semantically driven region detector. This represents the output prediction graph. The optimizer uses SGD, and the initial learning rate is set to 1×10. -4 The training employed a triangular warm-up strategy with 50 training cycles and a batch size of 16.

[0038] In this embodiment of the invention, the following experiments were conducted to further verify the feasibility and effectiveness of the method: In the experiment, publicly available camouflage target detection datasets were selected for training and testing. During the training phase, the COD10K and CAMO training sets were used for model training. During the testing phase, performance was evaluated on the CAMO, CHAMELEON, COD10K, and NC4K test sets. Four widely used evaluation metrics were used to quantitatively assess detection performance, including structural metrics (…). Enhanced alignment metrics ), weighted F-metric ( ) and mean absolute error (MAE); Specifically, structural metrics ( This is used to evaluate the similarity between the predicted mask and the real mask at the region structure level, reflecting the network structure's ability to characterize the overall shape and structural integrity of the camouflaged target; enhancing the alignment metric ( ) is used to simultaneously measure the alignment consistency between the predicted results and the ground truth annotations in terms of local details and global distribution, reflecting the overall consistency and boundary alignment quality of the detection results; weighted F-metric ( The method is used to assign weights to errors at different locations based on the combined precision and recall, which can more sensitively reflect the accuracy of the foreground anchor point region prediction and the false positives and false negatives. The mean absolute error (MAE) is used to measure the pixel-level average difference between the output prediction map and the ground truth annotation. The smaller the value, the smaller the prediction deviation, which can intuitively reflect the overall error level caused by false positives and false negatives. Tables 1 to 4 below show the quantitative comparison results of the method of this invention and the existing unsupervised camouflaged target detection methods in terms of various evaluation indicators on the CAMO test set, CHAMELEON test set, COD10K test set, and NC4K test set. Tables 1 to 4 are shown below: Table 1. Quantitative comparison of the method of this invention with existing unsupervised camouflaged target detection methods on the CAMO test set. Table 2. Quantitative comparison of the method of this invention with existing unsupervised camouflaged target detection methods on the CHAMELEON test set. Table 3. Quantitative comparison of the method of this invention with existing unsupervised camouflaged target detection methods on the COD10K test set. Table 4. Quantitative comparison of the method of this invention with existing unsupervised camouflaged target detection methods on the NC4K test set. As can be seen from the data in Tables 1 to 4 above, the method of the present invention has achieved better quantitative results on multiple test sets. It can effectively suppress the performance degradation caused by pseudo-supervision noise without manual annotation, improve the structural integrity and boundary accuracy of camouflaged target detection, and thus improve the stability and accuracy of camouflaged target detection.

[0039] In this embodiment of the invention, to intuitively demonstrate the detection effect of the method of the present invention, in addition to quantitative index comparison, this embodiment further performs visual comparison verification. For a given test image, the method of the present invention and existing unsupervised camouflage target detection methods are used respectively to obtain the corresponding camouflage target prediction results, which are then compared and analyzed with the ground truth (GT) annotations. Figure 6 The visualization of detection results for each method is presented in different scenario examples, showing the input image, ground truth annotations, and the output prediction maps of each comparison method. Figure 6 It is evident that existing unsupervised camouflage target detection methods are prone to problems such as incomplete target regions, false background activation, and blurred boundaries in camouflage scenes. In contrast, the method of this invention can generate more structurally complete target regions without manual annotation training, and maintain clearer contours and less background interference in slender structures, weak boundaries, and low-contrast regions. This demonstrates that the method of this invention has superior performance in terms of stability and accuracy in camouflage target detection.

[0040] In the description of this invention, the references to "one embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0041] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An unsupervised camouflage target detection method based on self-anchored progressive learning, characterized in that, Includes the following steps: Step S1: Obtain at least one image to be processed and input it into a pre-constructed self-supervised feature extraction network to extract multi-layer semantic features; Step S2: Based on the multi-layer semantic features, a semantically driven region detector is constructed, and cross-image contrastive learning is used to enhance the foreground and background separability to generate a region activation map. Step S3: Perform multi-scale fusion on the region activation map and use dual threshold filtering to obtain the foreground anchor region and background anchor region, and set the uncertain region between the thresholds as the ignored region to form region-level anchor label; Step S4: Based on the region-level anchor labels, a progressive contextual reasoning network is constructed, which integrates the multi-layer semantic features and performs cross-scale interactive reasoning. Under anchor region supervision, the structure of the camouflaged target is expanded and refined based on the foreground anchor region and the background anchor region to generate camouflaged target detection results.

2. The unsupervised camouflage target detection method of claim 1, wherein, In step S1, before inputting the image to be processed into the self-supervised feature extraction network, the method further includes: The image to be processed is subjected to size normalization processing, and the image to be processed is scaled to 320x320 pixels using the LANCZOS resampling method.

3. The unsupervised camouflaged target detection method according to claim 1, characterized in that, In step S1, the Vision Transformer self-supervised pre-trained model is used as the self-supervised feature extraction network, and the multi-layer semantic features include at least the last three layers of output features of the Vision Transformer self-supervised pre-trained model.

4. The unsupervised camouflaged target detection method according to claim 1, characterized in that, In step S2, the semantically driven region detector includes a feature transformation unit and parallel left, right, and summation branches. Semantic consistency enhancement processing is performed on the branch features extracted from each of the left, right, and summation branches. The semantic consistency enhancement processing includes: An attention map is generated based on the branch features. An uncertainty modulation map is generated by multiplying the attention map and its complement map element by element. The processed uncertainty modulation map is then fused with the branch features element by element to obtain the enhanced branch features.

5. The unsupervised camouflaged target detection method according to claim 4, characterized in that, In step S2, the enhanced branch features of each branch are concatenated along the channel dimension and then further convolved and the semantic consistency enhancement process is performed again to output the region feature map. Subsequently, a 3x3 convolution is performed on the region feature map through the decoupling head to obtain a single-channel activation map. Based on the single-channel activation map and its complement map, the region feature map is weighted and converged to obtain the foreground prototype features and the background prototype features.

6. The unsupervised camouflaged target detection method according to claim 5, characterized in that, In step S2, positive sample pairs are formed by cross-image contrastive learning using the foreground prototype features corresponding to different images or the background prototype features corresponding to different images, and negative sample pairs are formed using the foreground prototype features and the background prototype features of the same image, in order to enhance the foreground and background separability and generate the region activation map.

7. The unsupervised camouflaged target detection method according to claim 1, characterized in that, In step S3, multi-scale fusion uses multiple preset scales to process the region activation map to obtain the activation value of each region. Dual threshold filtering uses preset foreground threshold and background threshold. Regions with activation values ​​greater than the foreground threshold are marked as foreground anchor regions, regions with activation values ​​less than the background threshold are marked as background anchor regions, and regions with activation values ​​between the foreground threshold and the background threshold are marked as ignored regions.

8. The unsupervised camouflaged target detection method according to claim 1, characterized in that, In step S4, the anchor point region supervision condition is that supervision loss is calculated only for the foreground anchor point region and the background anchor point region, and supervision loss is not calculated for the ignored region.

9. The unsupervised camouflaged target detection method according to claim 1, characterized in that, The cross-scale interactive reasoning in step S4 includes: The feature channels are divided into multiple groups, and dilated convolutions with different dilation rates are used to aggregate the multi-layer semantic features within and between groups to output fused features. Then, adaptive average pooling is performed on the fused features to generate a global description, and the global description is used to generate attention weights through 1x1 convolution and nonlinear mapping. The attention weights are then used to weight the current semantic features and residual information respectively before fusion.

Images