Unsupervised camouflage target detection method based on confidence partition distillation and skeleton-induced detail reconstruction, electronic device and storage medium

Abstract

The application discloses an unsupervised camouflage target detection method based on confidence partition distillation and skeleton-induced detail reconstruction, an electronic device and a storage medium. The method comprises the following steps: extracting multi-level features of an image and generating initial pseudo-labels by using DINO; in the first stage, the initial pseudo-labels are dynamically fused with EMA prediction by a main teacher initialization module to generate dynamic pseudo-labels to train a teacher model, and student prediction is divided into a high-confidence area and a confusion area based on a confidence partition distillation module, and consistency distillation and confusion suppression constraints are implemented respectively; in the second stage, a predicted topological skeleton is extracted by using an active detail exploration module, and a detail refinement label is obtained by local temperature rescaling, and a student model is fine-tuned to restore boundary details; and finally, the student model is used to complete detection. The application can effectively suppress pseudo-label noise and restore target details without manual annotation, and significantly improves the camouflage target segmentation precision.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and specifically refers to an unsupervised camouflaged target detection method, electronic device, and storage medium based on confidence partition distillation and skeleton-induced detail reconstruction. Background Technology

[0002] Unsupervised camouflaged object detection aims to identify camouflaged objects in images that are highly similar to the surrounding background in color, texture, shape, and structure without relying on pixel-level manual annotation. Because the visual differences between camouflaged objects and the background are extremely subtle, this task requires models not only to discover hidden targets in complex environments but also to maintain the stability of target region localization and the integrity of boundary structures in the absence of real-world annotation supervision. Therefore, this technology has significant research value and practical implications for applications such as intelligent perception in complex open environments, ecological monitoring, and the discovery of hidden targets.

[0003] To suppress pseudo-label noise, existing techniques typically employ two improvement strategies: first, iteratively updating or dynamically adjusting pseudo-labels to allow the model's predictions to inversely optimize pseudo-label quality; second, using high-level semantic features or uncertainty estimation to screen and filter pseudo-labels, thereby reducing noise interference. While these methods can mitigate significant background noise to some extent, they still have significant shortcomings in camouflaged target detection scenarios. Specifically, on the one hand, because the boundaries of camouflaged targets are highly similar to the background, relying solely on overall confidence or a global threshold for pseudo-label screening makes it difficult to effectively distinguish confusing regions at the boundaries. This results in background regions similar to the target's appearance being continuously introduced into the training process, contaminating the supervision signal. On the other hand, while excessive noise filtering helps improve the purity of the main body region in pseudo-labels, it often simultaneously suppresses target boundaries, elongated structures, and low-contrast details, causing the supervision signal obtained by the model to be biased towards coarse-grained main body regions, making it difficult to recover the complete target outline.

[0004] Analysis reveals the following main shortcomings of existing unsupervised camouflaged target detection methods: (1) Relying solely on global confidence or fixed thresholds for pseudo-label screening cannot accurately distinguish the confusing areas near the boundary of the disguised target, resulting in the continuous propagation of background noise; (2) While excessive noise filtering strategies can improve the purity of the main region, they are prone to losing boundaries, elongated structures and low-contrast details, resulting in coarsening of the supervision signal and affecting the precision of the final segmentation. (3) The lack of a mechanism to implement differentiated learning constraints for different confidence regions fails to simultaneously take into account the stability of the main region and the integrity of the boundary details, thus limiting the performance of the model in extreme camouflage scenarios.

[0005] Therefore, in the field of unsupervised camouflaged target detection, how to effectively suppress pseudo-label noise without relying on manual annotation, and further utilize target structural information to guide the recovery of fine-grained boundaries and local details, remains a key technical problem that urgently needs to be solved. Specifically, there is a pressing need for a detection method that can fully utilize reliable supervision information in high-confidence regions, effectively suppress background noise in low-confidence confusion regions, and reconstruct target boundaries and local details through a structured detail guidance mechanism, thereby improving the target localization accuracy and segmentation integrity of the model in complex camouflaged scenarios. Summary of the Invention

[0006] The main objective of this invention is to provide an unsupervised method, electronic device, and storage medium for camouflaged target detection based on confidence partition distillation and skeleton-induced detail reconstruction. It employs a two-stage training strategy, from a "pseudo-label purification and confidence partition distillation stage" to a "skeleton-induced detail reconstruction stage." Without requiring manual pixel-level annotation, it first suppresses background noise propagation and stabilizes the high-confidence region representation of the camouflaged target through dynamic pseudo-labels and confidence partition distillation. Then, it restores the boundary and slender structural details of the camouflaged target through skeleton-induced detail region mining and local temperature rescaling, thereby achieving accurate segmentation of camouflaged targets highly similar to the background and generating a high-quality segmentation mask.

[0007] To achieve the above objectives, one solution of the present invention is: An unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction is proposed. This method employs a two-stage training strategy, from confidence partition distillation to skeleton-induced detail reconstruction, to train the model without manual annotation, achieving accurate segmentation of camouflaged targets highly similar to the background. The method includes the following steps: Step 1: Obtain an unlabeled training image sample set. Select training images from the training image sample set and use the pre-trained feature extraction backbone network DINO to extract features from the training images to obtain multi-level image features of the training images. Based on the last layer of features in the multi-level image features, use the BackgroundSeed method proposed by Siméoni et al. (Siméoni O, Sekkat C, Puy G, et al. Unsupervised Object Localization: Observing the Background To Discover Objects[C]. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2023: 3176-3186.) to construct the initial pseudo-labels of the training images.

[0008] Step 2: Enter the first stage, namely the pseudo-label purification and confidence partition distillation stage; based on the subject teacher initialization (STI) module, the initial pseudo-labels and the exponential moving average prediction results are dynamically fused to generate dynamic pseudo-labels, and the teacher model is trained using the dynamic pseudo-labels so that the teacher model outputs the camouflaged target prediction results with noise suppressed.

[0009] Step 3: In the first stage, the student model is further trained based on the Confidence Partition Distillation (SCD) module; according to the pixel confidence of the prediction results of the student model, the prediction region is divided into high-confidence regions and confused regions; for the high-confidence regions, teacher-student consistency distillation is performed to maintain reliable target region representation; for the confused regions, confusion suppression constraints are performed to reduce the propagation of erroneous knowledge under the condition of high similarity between foreground and background.

[0010] Step 4: Proceed to the second stage, namely the skeleton-induced detail reconstruction stage; Based on the Active Detail Exploration (ADE) module, perform topology mining and local temperature rescaling on the student model to obtain detail refinement labels; Use the detail refinement labels to fine-tune the student model and restore the prediction details of the student model.

[0011] Step 5: For the image to be detected, use the trained student model to make predictions, obtain the camouflage target probability map, and perform binarization on the camouflage target probability map to obtain the final camouflage target segmentation result.

[0012] In the preferred technical solution, step 2, the process of the main teacher initialization module training the teacher model specifically includes the following sub-steps: Step 2.1: Constructing a teacher model and its EMA copy Among them, the teacher model The DBA decoder proposed by Weiqi Yan et al. (Yan W, Chen L, Kou H, et al. UCOD-DPL: Unsupervised Camouflaged ObjectDetection via Dynamic Pseudo-label Learning[C]. Proceedings of the ComputerVision and Pattern Recognition Conference. 2025: 30365-30375.) is used, and its parameters are expressed as follows: EMA copy Use with Teacher Model The same DBA decoder, its parameters are expressed as ; During training, EMA parameters Instead of performing gradient backpropagation updates, teacher parameters are calculated. Update using the exponential moving average: ; in, , These represent the EMA parameters before and after the update, respectively. Indicates training steps; This represents the momentum coefficient, used to control the smoothness of the update process; The teacher parameters represent the training steps.

[0013] Step 2.2: Divide the multi-level image features The last layer of features Input teacher model and EMA copy In the middle, teachers' predictions were obtained. and EMA forecasts ; EMA forecast With the initial pseudo-label Perform linear mixing to obtain dynamic pseudo-labels. The calculation formulas are as follows: ; ; ; ; in, The number of layers representing multi-level image features; This represents the sigmoid activation function; Indicates the current training step count; This represents the total number of training steps.

[0014] Step 2.3: Using the aforementioned dynamic pseudo-tags As a supervisory signal, it is calculated to correlate with the teacher's prediction. Binary cross-entropy loss between : ; in, This represents the binary cross-entropy loss function.

[0015] In a preferred technical solution, in step 3, the confidence partitioning distillation module strengthens subject consistency in high-confidence regions and stimulates the self-discrimination ability of the student model in confused regions through a confidence partitioning strategy. Specifically, it includes the following sub-steps: Step 3.1: Construct a student model The student model Using the BiRefBlock decoder proposed by Peng Zheng et al. (PengZheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomy-mousimage segmentation. CAAI Artificial Intelligence Research, 2024.), multi-level image features are integrated. Input the student model In the middle, students' predictions were obtained. : .

[0016] Step 3.2: Use a preset deterministic threshold. For the student prediction Divide the regions according to confidence level, and define high-confidence regions. and confusing areas : ; ; in, This indicates an indicator function whose value is 1 when the expression inside is true, and 0 otherwise.

[0017] Step 3.3, targeting high-confidence areas First, construct the supervision objectives. The supervision target Predicted by the teacher With the student model Time-series stability prediction Linear mixing configuration: ; ; in, , These represent the time-series stable predictions before and after the update, respectively; This represents the time-series update coefficients of the student model prediction.

[0018] Step 3.4, targeting high-confidence areas The monitoring signal obtained in step 3.3 As a reliable monitoring signal, for student prediction Apply subject consistency constraints and calculate distillation loss in the high-confidence region. : ; in, Represents pixel coordinates; This indicates the high confidence mask at pixel coordinates. The possible values ​​of ; The segmentation loss is a combination of binary cross-entropy loss, intersection-union ratio loss, and structural similarity loss. Indicates pixel coordinates Student predictions; Indicate the monitoring objective; express__.

[0019] Step 3.5: Targeting the confused areas According to the preset parameters Select the highest and lowest predicted values ​​respectively. Pixels, to obtain the potential foreground set With potential background set ; for the set of potential prospects With potential background set Calculate the confusion suppression loss : ; in, This represents an operation that calculates the number of elements in a set. This indicates the operation for calculating the mean square error; This represents the student prediction confidence level of the selected pixel in the confusion region.

[0020] In a preferred technical solution, in step 4, the active detail exploration module mines the topological skeleton of the predicted map and utilizes an adaptive local temperature rescaling mechanism to actively sharpen blurred boundaries and generate fine-grained pseudo-labels while maintaining the main structure. Specifically, it includes the following sub-steps: Step 4.1: Process the student predictions using soft skeletonization. Extract the structural skeleton diagram : ; in, The operation of extracting the structural skeleton map was proposed by Suprosanna Shit et al. (Shit S, Paetzold JC, Sekuboyina A, et al. clDice-a novel topology-preserving loss function for tubular structure segmentation[C]. Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2021:16560-16569.). Based on the structural skeleton diagram The connectivity will predict the students Divide into several sets of disjoint connected components , to This represents the connected component that has been partitioned.

[0021] Step 4.2: For each connected component Calculate the local adaptive threshold of the current connected component. Based on the local adaptive threshold Temperature rescaling is applied to the current connected component to obtain detailed labels. : ; ; ; in, This indicates a temperature rescaling operation; This refers to the logit function; This indicates the preset temperature scaling parameters; Represents the binary mask corresponding to the connected component; This indicates element-wise multiplication.

[0022] Step 4.3: Refine the labels according to the details mentioned above. As a supervisory signal, its correlation with student predictions is calculated. Detailed exploration loss : .

[0023] The second solution of the present invention is: An electronic device includes a processor, a memory, and an application program; the application program is stored in the memory and configured to be executed by the processor using the unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction.

[0024] The third solution of the present invention is: A computer-readable storage medium storing a computer program; when executed in the computer, the computer program causes the computer to perform the unsupervised camouflage target detection method based on confidence partition distillation and skeleton-induced detail reconstruction.

[0025] After adopting the above technical solution, the present invention has the following technical effects: (I) This invention proposes an unsupervised camouflaged target detection method for suppressing pseudo-label noise. By combining confidence partition distillation with structured detail reconstruction, differential learning constraints are applied to prediction regions with different reliability levels, thereby alleviating the problems of easy propagation of background noise and easy loss of boundary details in existing methods.

[0026] (II) The present invention constructs a subject teacher initialization (STI) module, which uses a pseudo-label fusion mechanism that changes dynamically with the training process to adaptively combine the initial pseudo-labels with the model’s historical predictions, gradually reducing the background noise interference in the initial pseudo-labels and providing more reliable subject region supervision signals for subsequent student model learning.

[0027] (III) The present invention designs a subject consistency distillation (SCD) module mechanism, which divides the prediction region into a high confidence region and a confusion region, and implements teacher-student knowledge transfer and self-supervised noise suppression respectively, so as to reduce the propagation of erroneous knowledge under the condition that the foreground and background are highly similar.

[0028] (IV) This invention constructs an Active Detail Exploration (ADE) module, which uses the topological skeleton information in the prediction results to locate potential boundaries and slender structural regions, and combines a local temperature rescaling mechanism to generate fine-grained pseudo-labels, thereby enhancing the recovery capability of target boundaries, slender structures and local details without the need for manual ground truth labels.

[0029] (V) This invention can simultaneously improve the localization stability and boundary structure integrity of the camouflaged target's main body region under unsupervised training conditions. Experimental results show that this invention performs excellently on multiple camouflaged target detection (COD) benchmark sets. Compared with existing unsupervised methods, this invention achieves significant improvements in key indicators such as mean absolute error (MAE), S-Measure, E-Measure, and F-Measure, and reaches or even surpasses the performance level of some fully supervised methods. Attached Figure Description

[0030] Figure 1 This is a flowchart of a specific embodiment of the present invention.

[0031] Figure 2 These are experimental results from specific embodiments of the present invention. Detailed Implementation

[0032] To further explain the technical solution of the present invention, the present invention will be described in detail below through specific embodiments.

[0033] refer to Figure 1 As shown, this invention discloses an unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction. Based on a two-stage training strategy from confidence partition distillation to skeleton-induced detail reconstruction, the model is trained without manual annotation, achieving accurate segmentation of camouflaged targets highly similar to the background. The method includes the following steps: Step 1: Obtain an unlabeled training image sample set. Select training images from the training image sample set and use the pre-trained feature extraction backbone network DINO to extract features from the training images to obtain multi-level image features of the training images. Based on the last layer of features in the multi-level image features, the Background Seed method proposed by Siméoni et al. (Siméoni O, Sekkat C, Puy G, et al. Unsupervised Object Localization: Observing the Background ToDiscover Objects[C]. Proceedings of the IEEE / CVF Conference on ComputerVision and Pattern Recognition. 2023: 3176-3186.) is used to construct the initial pseudo-labels of the training images.

[0034] Step 2: Enter the first stage, namely the pseudo-label purification and confidence partition distillation stage; based on the subject teacher initialization (STI) module, the initial pseudo-labels and the exponential moving average prediction results are dynamically fused to generate dynamic pseudo-labels, and the teacher model is trained using dynamic pseudo-labels so that the teacher model outputs the camouflaged target prediction results with noise suppressed.

[0035] Furthermore, in step 2 above, the process of training the teacher model using the main teacher initialization module specifically includes the following sub-steps: Step 2.1: Constructing a teacher model and its EMA copy Among them, the teacher model The DBA decoder proposed by Weiqi Yan et al. (Yan W, Chen L, Kou H, et al. UCOD-DPL: Unsupervised Camouflaged ObjectDetection via Dynamic Pseudo-label Learning[C]. Proceedings of the ComputerVision and Pattern Recognition Conference. 2025: 30365-30375.) is used, and its parameters are expressed as follows: EMA copy Use with Teacher Model The same DBA decoder, its parameters are expressed as ; During training, EMA parameters Instead of performing gradient backpropagation updates, teacher parameters are calculated. Update using the exponential moving average: ; in, , These represent the EMA parameters before and after the update, respectively. Indicates training steps; This represents the momentum coefficient, used to control the smoothness of the update process; The teacher parameters represent the training steps.

[0036] Step 2.2: Divide the multi-level image features The last layer of features Input teacher model and EMA copy In the middle, teachers' predictions were obtained. and EMA forecasts ; EMA forecast With the initial pseudo-label Perform linear mixing to obtain dynamic pseudo-labels. The calculation formulas are as follows: ; ; ; ; in, The number of layers representing multi-level image features; This represents the sigmoid activation function; Indicates the current training step count; This represents the total number of training steps.

[0037] Step 2.3: Use dynamic pseudo-tags As a supervisory signal, its correlation with teacher predictions is calculated. Binary cross-entropy loss between : ; in, This represents the binary cross-entropy loss function.

[0038] Step 3: In the first stage, the student model is further trained based on the Confidence Partition Distillation (SCD) module; according to the pixel confidence of the prediction results of the student model, the predicted region is divided into high confidence region and confused region; for high confidence region, teacher-student consistency distillation is performed to maintain reliable target region representation; for confused region, confusion suppression constraint is performed to reduce the propagation of erroneous knowledge under the condition of high similarity between foreground and background.

[0039] Furthermore, in step 3 above, the confidence partitioning distillation module strengthens subject consistency in high-confidence regions and stimulates the self-discrimination ability of the student model in confusing regions through a confidence partitioning strategy. Specifically, it includes the following sub-steps: Step 3.1: Construct a student model Student Model Using the BiRefBlock decoder proposed by Peng Zheng et al. (Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomy-mousimage segmentation. CAAI Artificial Intelligence Research, 2024.), multi-level image features are integrated. Input student model In the middle, students' predictions were obtained. : .

[0040] Step 3.2: Use a preset deterministic threshold. Predictions for students Divide the regions according to confidence level, and define high-confidence regions. and confusing areas : ; ; in, This indicates an indicator function whose value is 1 when the expression inside is true, and 0 otherwise.

[0041] Step 3.3, targeting high-confidence areas First, construct the supervision objectives. Supervision target Teacher prediction With student model Time-series stability prediction Linear mixing configuration: ; ; in, , These represent the time-series stable predictions before and after the update, respectively; This represents the time-series update coefficients of the student model prediction.

[0042] Step 3.4, targeting high-confidence areas The monitoring signal obtained in step 3.3 As a reliable monitoring signal, for student prediction Apply subject consistency constraints and calculate distillation loss in the high-confidence region. : ; in, Represents pixel coordinates; This indicates the high confidence mask at pixel coordinates. The possible values ​​of ; The segmentation loss is a combination of binary cross-entropy loss, intersection-union ratio loss, and structural similarity loss. Indicates pixel coordinates Student predictions; Indicate the monitoring objective; express__.

[0043] Step 3.5: Targeting the confused areas According to the preset parameters Select the highest and lowest predicted values ​​respectively. Pixels, to obtain the potential foreground set With potential background set ; for the set of potential prospects With potential background set Calculate the confusion suppression loss : ; in, This represents an operation that calculates the number of elements in a set. This indicates the operation for calculating the mean square error; This represents the student prediction confidence level of the selected pixel in the confusion region.

[0044] Step 4: Enter the second stage, namely the skeleton-induced detail reconstruction stage; based on the Active Detail Exploration (ADE) module, perform topology mining and local temperature rescaling on the student model to obtain detail refinement labels; use the detail refinement labels to fine-tune the student model and restore the predictive details of the student model.

[0045] Furthermore, in step 4 above, the active detail exploration module mines the topological skeleton of the predicted map and utilizes an adaptive local temperature rescaling mechanism to actively sharpen blurred boundaries and generate fine-grained pseudo-labels while maintaining the main structure. Specifically, this includes the following sub-steps: Step 4.1: Process student predictions using soft skeletonization. Extract the structural skeleton diagram : ; in, The operation of extracting the structural skeleton map was proposed by Suprosanna Shit et al. (Shit S, Paetzold JC, Sekuboyina A, et al. clDice-a novel topology-preserving loss function for tubular structure segmentation[C]. Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2021:16560-16569.). Based on structural skeleton diagram Connectivity will help students predict Divide into several sets of disjoint connected components , to This represents the connected component that has been partitioned.

[0046] Step 4.2: For each connected component Calculate the local adaptive threshold of the current connected component. Based on local adaptive threshold Temperature rescaling is applied to the current connected component to obtain detailed labels. : ; ; ; in, This indicates a temperature rescaling operation; This refers to the logit function; This indicates the preset temperature scaling parameters; Represents the binary mask corresponding to the connected component; This indicates element-wise multiplication.

[0047] Step 4.3: Refine the tags with details As a supervisory signal, its correlation with student predictions is calculated. Detailed exploration loss : .

[0048] Step 5: For the image to be detected, use the trained student model to make predictions, obtain the camouflage target probability map, and perform binarization on the camouflage target probability map to obtain the final camouflage target segmentation result.

[0049] The technical effects of this invention are demonstrated below through experimental data: This invention follows the common setup in the field of camouflaged target detection, using a combination of the CAMO training set and the COD10K training set (excluding images that do not contain camouflaged targets) as the training set. No corresponding labels are used during the training process.

[0050] For the selection of test sets, this invention conducts tests on four datasets: CHAMELEON (containing 76 test images), CAMO-Test (containing 250 test images), COD10K-Test (containing 2026 test images), and NC4K (containing 4121 test images). All test sets have pixel-level manual annotations.

[0051] Experimental results are as follows Figure 2 As shown, experimental results demonstrate that the present invention outperforms existing semi-supervised camouflage target detection models on multiple metrics and surpasses some existing fully supervised camouflage target detection models. Compared with the current state-of-the-art semi-supervised camouflage target detection models, SEED achieves an average improvement of 23.02% in MAE and an average improvement of 2.47% in S-measure, proving the effectiveness of the model proposed in this invention.

[0052] Extensive experiments conducted on the above datasets demonstrate that the proposed method achieves significant and consistent performance improvements compared to other competing methods.

[0053] The present invention also discloses an electronic device, including a processor, a memory, and an application program; the application program is stored in the memory and configured to be executed by the processor using the methods described above.

[0054] The present invention also discloses a computer-readable storage medium storing a computer program; when the computer program is executed in a computer, it causes the computer to perform the above-described method. The storage medium may include various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0055] The above embodiments and figures are not intended to limit the product form and style of the present invention. Any appropriate changes or modifications made by those skilled in the art should be considered as not departing from the patent scope of the present invention.

Claims

1. An unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction, characterized in that, Includes the following steps: Step 1: Obtain an unlabeled training image sample set, select training images from the training image sample set, and use the pre-trained feature extraction backbone network DINO to extract features from the training images to obtain multi-level image features of the training images; based on the last layer of features in the multi-level image features, construct the initial pseudo-label of the training images. Step 2: Enter the first stage, namely the pseudo-label purification and confidence partition distillation stage; Based on the main teacher initialization module, the initial pseudo-labels are dynamically fused with the exponential moving average prediction results to generate dynamic pseudo-labels, and the teacher model is trained using the dynamic pseudo-labels so that the teacher model outputs a disguised target prediction result with noise suppressed. Step 3: In the first stage, a student model is trained based on the confidence partition distillation module; according to the pixel confidence of the prediction results of the student model, the prediction region is divided into high-confidence regions and confused regions; for the high-confidence regions, teacher-student consistency distillation is performed to maintain reliable target region representation; for the confused regions, confusion suppression constraints are performed to reduce the propagation of erroneous knowledge under the condition of high similarity between foreground and background. Step 4: Proceed to the second stage, namely the skeleton-induced detail reconstruction stage; Based on the active detail exploration module, perform topology mining and local temperature rescaling on the student model to obtain detail refinement labels; Use the detail refinement labels to fine-tune the student model and restore the predicted details of the student model; Step 5: For the image to be detected, use the trained student model to make predictions, obtain the camouflage target probability map, and perform binarization on the camouflage target probability map to obtain the final camouflage target segmentation result.

2. The unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2.1: Constructing a teacher model and its EMA copy Among them, the teacher model Using the DBA decoder, its parameters are expressed as follows: EMA copy Use with Teacher Model The same DBA decoder, its parameters are expressed as During the training process, teacher parameters are calculated. The exponential moving average is used to update the EMA parameter. : ； in, , These represent the EMA parameters before and after the update, respectively. Indicates training steps; This represents the momentum coefficient, used to control the smoothness of the update process; Teacher parameters indicating training steps; Step 2.2: Divide the multi-level image features The last layer of features Input teacher model and EMA copy In the middle, teachers' predictions were obtained. and EMA forecasts ; EMA forecast With the initial pseudo-label Perform linear mixing to obtain dynamic pseudo-labels. The calculation formulas are as follows: ； ； ； ； in, The number of layers representing multi-level image features; This represents the sigmoid activation function; Indicates the current training step count; This represents the total number of training steps; Step 2.3: Using the aforementioned dynamic pseudo-tags As a supervisory signal, it is calculated to correlate with the teacher's prediction. Binary cross-entropy loss between : ； in, This represents the binary cross-entropy loss function.

3. The unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction according to claim 2, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Construct a student model The student model Use the BiRefBlock decoder to extract multi-level image features. Input the student model In the middle, students' predictions were obtained. : ； Step 3.2: Use a preset deterministic threshold. For the student prediction Divide the regions according to confidence level, and define high-confidence regions. and confusing areas : ； ； in, This indicates an indicator function; its value is 1 if the expression inside it is true, and 0 otherwise. Step 3.3, targeting high-confidence areas First, construct the supervision objectives. The supervision target Predicted by the teacher With the student model Time-series stability prediction Linear mixing configuration: ； ； in, , These represent the time-series stable predictions before and after the update, respectively; This represents the time-series update coefficients of the student model predictions; Step 3.4, targeting high-confidence areas The monitoring signal obtained in step 3.3 As a reliable monitoring signal, for student prediction Apply subject consistency constraints and calculate distillation loss in the high-confidence region. : ； in, Represents pixel coordinates; This indicates the high confidence mask at pixel coordinates. The possible values ​​of ; The segmentation loss is a combination of binary cross-entropy loss, intersection-union ratio loss, and structural similarity loss. Indicates pixel coordinates Student predictions; Indicate the monitoring objective; express__; Step 3.5: Targeting the confused areas According to the preset parameters Select the highest and lowest predicted values ​​respectively. Pixels, to obtain the potential foreground set With potential background set ; for the set of potential prospects With potential background set Calculate the confusion suppression loss : ； in, This represents an operation that calculates the number of elements in a set. This indicates the operation for calculating the mean square error; This represents the student prediction confidence level of the selected pixel in the confusion region.

4. The unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction according to claim 3, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Process the student predictions using soft skeletonization. Extract the structural skeleton diagram : ； in, This indicates the operation of extracting the structural skeleton diagram; Based on the structural skeleton diagram The connectivity will predict the students Divide into several sets of disjoint connected components , to Represents the partitioned connected components; Step 4.2: For each connected component Calculate the local adaptive threshold of the current connected component. Based on the local adaptive threshold Temperature rescaling is applied to the current connected component to obtain detailed labels. : ； ； ； in, This indicates a temperature rescaling operation; This refers to the logit function; This indicates the preset temperature scaling parameters; Represents the binary mask corresponding to the connected component; This represents element-wise multiplication; Step 4.3: Refine the labels according to the details mentioned above. As a supervisory signal, its correlation with student predictions is calculated. Detailed exploration loss : 。 5. An electronic device, characterized in that, It includes a processor, a memory, and an application program; the application program is stored in the memory and configured to be executed by the processor as described in any one of claims 1 to 4, for the unsupervised camouflaged target detection method based on confidence partition distillation and skeleton-induced detail reconstruction.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program; when the computer program is executed in the computer, it causes the computer to perform the unsupervised camouflage target detection method based on confidence partition distillation and skeleton-induced detail reconstruction according to any one of claims 1 to 4.

Images