Weak supervision camouflage target detection method and system based on graffiti labeling and depth information assistance

Abstract

The invention discloses a weak supervision camouflage target detection method and a weak supervision camouflage target detection system based on graffiti labeling and depth information assistance, and designs a cooperative training and feature fusion mechanism through graffiti labeling with extremely low cost in combination with complementary information of RGB and depth images. Through multi-modal information fusion and a pseudo label optimization mechanism, on the premise of only needing sparse graffiti labeling, high-precision camouflage target detection is realized while the data labeling burden is remarkably reduced, scene understanding and detail retention capability which are not inferior to that of a fully supervised model are realized, and the detection accuracy is improved. And a practical solution is provided for low-cost and high-precision camouflage target detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and in particular relates to a method for detecting camouflaged targets. Background Technology

[0002] Camouflage target detection aims to identify objects that blend seamlessly with the background and has significant application value across various fields. On the one hand, current mainstream methods rely on pixel-level fully supervised learning, requiring a large amount of finely labeled data, resulting in extremely high annotation costs and heavy dependence on the professional level of annotators, thus limiting the technology's widespread adoption and application expansion. For example, Chinese patent CN202511209344.8 discloses a camouflage target detection method combining attention and pyramid decoding, but it relies on highly detailed pixel-level labeled data. On the other hand, existing research mostly utilizes only RGB images, and in extreme cases where the target perfectly matches the background color and texture, detection performance encounters bottlenecks. To reduce annotation costs, weakly supervised learning has become a research direction.

[0003] However, existing weakly supervised methods face problems such as sparse supervision signals, low quality of generated pseudo-labels, and blurred boundaries in camouflaged scenarios, resulting in insufficient detection accuracy. Furthermore, how to effectively utilize multimodal information such as depth to compensate for the deficiencies of RGB information, and how to achieve effective fusion of multimodal information within a weakly supervised framework, remain pressing technical challenges that need to be addressed.

[0004] Therefore, there is an urgent need for a new method that can achieve high-precision camouflage target detection with extremely low annotation costs and full utilization of multimodal information. Summary of the Invention

[0005] Purpose of the invention: This invention aims to overcome the shortcomings of the prior art and provide a weakly supervised camouflage target detection method and system based on graffiti annotation and depth information assistance. By using extremely low-cost graffiti annotation and combining complementary information from RGB and depth images, a collaborative training and feature fusion mechanism is designed to achieve high-precision camouflage target detection while significantly reducing the data annotation burden.

[0006] Technical solution:

[0007] This invention proposes a weakly supervised camouflage target detection method based on graffiti annotation and depth information assistance, comprising:

[0008] Step 1: Obtain an RGB image and a depth image containing the camouflaged target, and add graffiti annotations to the RGB image to obtain a graffiti-annotated image;

[0009] Step 2: Process the RGB image and the depth image separately using the superpixel segmentation algorithm, and combine them with the doodle annotation image to generate initial RGB pseudo-labels and initial depth pseudo-labels;

[0010] Step 3: Use the two original PVTv2 networks as single-modal prediction modules to process the RGB image and the depth image respectively, and generate predicted RGB pseudo-labels and predicted depth pseudo-labels; use a pixel adaptive thinning algorithm to aggregate the predicted RGB pseudo-labels and predicted depth pseudo-labels to obtain aggregated pseudo-labels;

[0011] Step 4: Construct a prediction model, including an encoder for extracting image features, an aggregation module for fusing features, and a decoder for outputting a prediction map. Design the composite loss function of the prediction model and train the prediction model end-to-end based on a public dataset.

[0012] The composite loss function includes: graffiti supervision loss based on the graffiti-annotated image, single-modal prediction loss based on the initial RGB pseudo-label and the initial depth pseudo-label, and aggregate supervision loss based on the aggregated pseudo-label;

[0013] Step 5: Input the RGB image and depth image to be detected into the trained prediction model, and output the predicted image of the camouflaged target.

[0014] Furthermore, the process of using a superpixel segmentation algorithm to process the RGB image and the depth image includes:

[0015] The superpixel segmentation algorithm is applied to the RGB image to obtain several superpixel regions of the RGB image. The intersection of the superpixel regions of the graffiti annotation image and the RGB image is taken to obtain the initial RGB pseudo-label.

[0016] A superpixel segmentation algorithm is applied to the depth image to obtain several superpixel regions of the depth image and the superpixel region of the depth image. The intersection of the superpixel regions of the graffiti annotation image and the depth image is taken to obtain the initial depth pseudo-label.

[0017] Furthermore, when the single-modal prediction module processes the RGB image, it takes the RGB image as input, extracts several RGB image features layer by layer from low to high, uses the initial RGB pseudo-label obtained in step 2 as supervision, decodes the extracted highest-level RGB image features, and generates predicted RGB pseudo-labels.

[0018] When the single-modal prediction module processes the depth image, it takes the depth image as input, extracts several depth image features layer by layer from low to high, uses the initial depth pseudo-label obtained in step 2 as supervision, decodes the extracted highest layer depth image features, and generates predicted depth pseudo-labels.

[0019] Furthermore, the pixel adaptive thinning algorithm includes: taking the predicted RGB pseudo-label and the predicted depth pseudo-label as input, guided by the color and spatial affinity of the RGB image, performing iterative optimization and weighted averaging, and outputting aggregated pseudo-labels.

[0020] Furthermore, the encoder of the prediction model is a PVTv2 network that extracts several layers of RGB image features from the RGB image input to the prediction model and several layers of depth image features from the depth image input to the prediction model. The number of layers of the RGB image features is the same as the number of layers of the depth image features.

[0021] Furthermore, the aggregation module of the prediction model includes a depthwise separable pyramid pooling module and an attention fusion module. The depthwise separable pyramid pooling module performs context enhancement on the highest-level features extracted by the encoder, and the attention fusion module fuses the RGB image features and depth image features of the same layer into a cross-modal fusion feature. The decoder of the prediction model aggregates the fusion features generated by the aggregation module from top to bottom and outputs a prediction map.

[0022] Furthermore, the graffiti supervision loss includes: obtaining a partial cross-entropy loss by supervising the prediction map of the prediction model through graffiti-annotated images. Local coherence loss and structural consistency loss Summing up the losses from graffiti supervision .

[0023] Furthermore, the single-modal prediction loss includes: based on the initial RGB pseudo-labels, supervising the predicted RGB pseudo-labels output by the single-modal prediction module to obtain the cross-entropy loss. Based on the initial depth pseudo-labels, the predicted depth pseudo-labels output by the single-modal prediction module are supervised, and the resulting cross-entropy loss is calculated. .

[0024] Furthermore, the aggregated supervision loss includes: supervising the prediction graph of the prediction model based on the aggregated pseudo-labels described in step 4 to obtain a weighted cross-entropy loss. and weighted IoU loss The summation yields the aggregated supervised loss. .

[0025] This invention also proposes a weakly supervised camouflage target detection system based on graffiti annotation and depth information assistance, comprising:

[0026] The data preparation module is used to acquire RGB images and depth images containing the camouflaged target, and to add graffiti annotations to the RGB images to obtain graffiti-annotated images;

[0027] The initial pseudo-label module is used to process the RGB image and the depth image respectively using a superpixel segmentation algorithm, and combine the graffiti annotation image to generate initial RGB pseudo-labels and initial depth pseudo-labels;

[0028] The aggregation pseudo-label module is used to process the RGB image and the depth image respectively using two original PVTv2 networks as single-modal prediction modules to generate predicted RGB pseudo-labels and predicted depth pseudo-labels; the predicted RGB pseudo-labels and predicted depth pseudo-labels are aggregated using a pixel adaptive thinning algorithm to obtain aggregated pseudo-labels.

[0029] The prediction model includes an encoder for extracting image features, an aggregation module for fusing features, and a decoder for outputting a prediction map. A composite loss function for the prediction model is designed, and the prediction model is trained end-to-end based on a public dataset. The composite loss function includes: graffiti supervision loss based on the graffiti-annotated image, single-modal prediction loss based on the initial RGB pseudo-labels and the initial depth pseudo-labels, and aggregation supervision loss based on the aggregated pseudo-labels.

[0030] The output module is used to process the RGB image and depth image to be detected through the prediction model and output a predicted map of the camouflaged target.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0032] First, the annotation cost is extremely low, requiring only sparse doodle annotations, without the need for expensive and time-consuming pixel-level fine annotations, which greatly reduces the threshold and cost of data preparation.

[0033] Secondly, it achieves high detection accuracy by incorporating depth information to compensate for the shortcomings of RGB modality in terms of 3D structure and geometric information. An innovative multi-level context aggregation module, including a depth-separable pyramid pooling module and an attention fusion module, effectively fuses and enhances bimodal features, enabling the model to maintain high detection accuracy even when the target and background are highly similar. Experiments show that, under weak supervision, this method's performance can match or even surpass some classic fully supervised methods.

[0034] Third, it has strong training robustness. The proposed "cooperative pseudo-label generation-prediction optimization-aggregation" framework can gradually generate high-quality aggregated pseudo-labels from sparse scribbles and noisy initial pseudo-labels, providing stable and reliable supervision signals for the training of prediction models and solving the problem of high noise in weak supervision learning. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention;

[0036] Figure 2 A flowchart for generating pseudo-labels using a superpixel segmentation algorithm;

[0037] Figure 3 This is a structural diagram of a single-mode prediction module;

[0038] Figure 4 This is a structural diagram of the prediction model;

[0039] Figure 5 This is a comparison chart of the detection results of this method and other methods. Detailed Implementation

[0040] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. Figure 1 The diagram shows a flowchart of a weakly supervised camouflage target detection method based on graffiti annotation and depth information assistance according to the present invention, which specifically includes the following steps:

[0041] Step 1: Data preparation and input.

[0042] RGB images containing camouflaged targets and their corresponding depth images are acquired, and the RGB images are annotated with extremely sparse doodles as initial weak supervision signals. In this embodiment, four publicly available mainstream camouflaged target datasets are collected and organized, including CAMO, CHAMELEON, COD10K, and NC4K. The S-COD public training set (download address: https: / / github.com / dddraxxx / Weakly-Supervised-Camouflaged-Object-Detection-with-Scribble-Annotations), which includes 1000 CAMO training images and 3040 COD10K training images, is selected. Random dots or short lines are annotated with doodles on the RGB images in the training set. An RGB image, its corresponding depth image, and its corresponding doodle annotation image are considered as a pair of training samples.

[0043] Step 2: Initial pseudo-tag generation.

[0044] For each pair of training samples, collaborative pseudo-label generation is performed first. A superpixel segmentation algorithm based on Simple Linear Iterative Clustering (SLIC) is used to generate superpixel regions for both the RGB and depth images. This superpixel segmentation algorithm is based on the research of Radhakrishna Achanta (SLIC Superpixels Compared to State-of-the-Art Superpixel Methods, doi: 10.1109 / TPAMI.2012.120.). The superpixel regions obtained by the superpixel segmentation algorithm in this embodiment are as follows: Figure 2 As shown, it's worth noting that a superpixel region is a collection of points and does not inherently possess color attributes. To make the segmentation results more visually appealing, Figure 2 Different colors are used to distinguish the positions of different superpixel regions.

[0045] The intersection of the segmented RGB image superpixel regions and the graffiti annotation image is taken, and the complete overlapping regions are preserved. The overall output forms the initial RGB pseudo-label G. r Similarly, the intersection of the segmented depth image superpixel regions and the graffiti annotation map is used to form the depth pseudo-label G. d .

[0046] Step 3: Predicting pseudo-label generation and optimizing aggregation.

[0047] Use the initial RGB pseudo-label G generated in step 2 respectively r and initial depth pseudo-label G d It supervises two independent unimodal prediction modules to generate smoother, higher-quality predicted RGB pseudo-labels and predicted depth pseudo-labels.

[0048] The single-modal prediction module is a basic PVTv2 network, such as... Figure 3 The diagram shows the structure of the single-modal prediction module in this invention. First, the RGB image is processed using the original PVTv2 network to extract hierarchical features, resulting in R1, R2, R3, and R4. The same operation is performed on the depth image, yielding D1, D2, D3, and D4. R4 and D4 are processed through continuous convolution and upsampling operations to smooth the prediction results. Continuous upsampling after convolution provides more supervisory information. The auxiliary supervisory loss stage uses a lightweight convolutional network, with the initial RGB pseudo-label G obtained in step 2 as the input. r and initial depth pseudo-label G d For supervision, the R4 and D4 features are decoded to generate smoothed RGB prediction pseudo-labels P. r And depth prediction pseudo-label Pd .

[0049] The pixel-adaptive refinement (PAR) module is used to refine the two predicted pseudo-labels P mentioned above. r and P d Iterative optimization and aggregation are performed to generate a high-quality, unified aggregated pseudo-label. The PAR module here references Lixiang Ru's research (Learning Affinity from Attention: End-to-End Weakly-Supervised Semantic Segmentation with Transformers, arXiv:2203.02664 [cs.CV]), specifically: P... r and P d The input to the PAR module is guided by the color and spatial affinity of the RGB image to perform iterative optimization and weighted averaging, resulting in a high-quality final aggregated pseudo-label Y.

[0050] The PAR module is used for aggregation and optimization of the prediction map, mainly because it can utilize the image's own visual features such as color, texture, and spatial structure to guide and correct high-level semantic predictions, achieving spatial consistency and boundary alignment of the prediction results. The core advantages of this method are threefold: First, it requires no additional annotation, mining the image's inherent structural priors through self-supervised learning; second, it employs multi-scale dilated neighborhood aggregation, capturing both local details and modeling long-range dependencies, adapting to camouflaged targets of different sizes; and finally, it uses an iterative optimization mechanism to gradually propagate high-confidence predictions, suppressing noise and ultimately generating pseudo-labels with clear boundaries and uniform internal structure.

[0051] Step 4: Construct a prediction model to achieve multi-scale feature extraction and fusion.

[0052] like Figure 4 The diagram shows the structure of the prediction model, which includes an encoder, a multi-level context aggregation module, and a decoder. The RGB image and depth image from step 1 are used as inputs to the prediction model.

[0053] First, an encoder is constructed to extract multi-level features from both the RGB and depth images. In this embodiment, the encoder chosen is PVTv2, whose global attention mechanism and multi-scale pyramid structure are particularly suitable for spoofing detection scenarios requiring strong contextual understanding. Theoretically, this encoder can be replaced with other encoders capable of image feature extraction.

[0054] Let the input RGB image be Depth image is Where H is the image height, W is the image width, and C is the image height. d This represents the number of channels in the depth image. PVTv2 outputs four levels of features: {R1-R4} and {D1-D4}, where {R1-R4} represents the features of the RGB image, and {D1-D4} represents the features of the depth image, specifically as follows:

[0055]

[0056]

[0057] in, , Represents feature hierarchy, spatial resolution Decreasing layer by layer, number of channels Increasing layer by layer. Encoder rgb For encoder calculation, F rgb These are the feature values. Taking the first layer of features (R1 and D1) as an example, R1 is a single-scale feature value, belonging to R... H1*W1*C1 The spatial resolution of R1D1 is (H1,W1)=(H / 4,W / 4).

[0058] Secondly, a multi-level context aggregation module is designed, comprising two parts: a Depthically Separable Pyramid Pooling (DSPP) module and an Attention Fusion (AF) module. The DSPP module is used to enhance high-level features with multi-scale contextual information, capturing target cues within different receptive fields. The AF module is used to fuse RGB and depth features layer by layer. This module adaptively weights and integrates bimodal information through the synergistic effect of channel attention and spatial attention, suppressing redundant noise and highlighting complementary and effective features.

[0059] The AF module also includes concat concatenation, 1*1 convolutional kernels, 3*3 convolutional kernels, channel attention, and spatial attention, with the following structure: Figure 4 As shown in the magnified section on the right. Channel attention (CALayer) enhances the discriminative power of feature representation, while spatial attention (SpatialAtt) focuses on potential target regions. The coordination of these two approaches can dynamically suppress redundant or noisy information and exhibit strong robustness when there are conflicts in cross-modal information. This achieves adaptive complementary fusion of texture information from the RGB modality and geometric and distance information from the depth modality.

[0060] In the aggregation module, firstly, the highest-level features R4 and D4 need to be enhanced using the DSPP module for multi-scale contextualization, and then the RGB and depth features of each level are fused using the AF module. The specific process is as follows:

[0061] Step 4.1: Perform Global Average Pooling (GAP), which averages the pixels across all spatial locations of the corresponding channel. The resulting feature map of size [B, C, H, W] is transformed into [B, C, 1, 1] after GAP, where B is the batch size and C is the number of channels. The formula for Global Average Pooling is as follows:

[0062]

[0063] Where b is the batch index and c is the channel index. Then, the GAP results are upsampled using a convolution kernel, calculated as follows:

[0064]

[0065] Where G is the global context branch, the enhanced high-level features are obtained through concatenated convolutions of multiple deep separable dilated convolution branches and the global context branch. , The calculation process is as follows:

[0066]

[0067]

[0068]

[0069]

[0070] in For point convolution branches, For depth-separable dilated convolution branches, The DSPP module can capture multi-scale contextual information around the target, which is crucial for distinguishing camouflaged targets that are highly similar to the background.

[0071] Step 4.2: The RGB and depth features at each level are fused through an attention mechanism to generate cross-modal representations. The calculation process is as follows:

[0072] First, concatenate the corresponding RGB features (R1-R4) with the depth features (D1-D4):

[0073]

[0074]

[0075] In the above formula, concat is a concatenation operation that connects multiple tensors along a specified dimension to form a new tensor. This is the splicing result of level l.

[0076] Next, the features are initially fused using convolution:

[0077]

[0078] In the above formula, Conv represents the convolution kernel and the convolution operation. This is the fusion result for level l.

[0079] Step 4.3: Perform global average pooling on the fusion result, then multiply it by weights generated through a multilayer perceptron and the sigmoid function.

[0080]

[0081] in The sigmoid activation function is used, MLP is a multilayer perceptron, and GAP is global average pooling. The output of the channel attention is subjected to average pooling and max pooling along the channel dimension to generate two feature maps. These feature maps are concatenated along the channel dimension and then fused using a 7x7 convolutional kernel. Finally, the spatial attention weights are generated using the sigmoid activation function, calculated as follows:

[0082]

[0083] in It is the Sigmoid activation function. The kernel is 7*7, and AvgPool and MaxPool are average pooling and max pooling, respectively.

[0084] After steps 4.1 to 4.3 above, the multi-level context aggregation module has obtained multi-scale features {RD4} to {RD1}.

[0085] Finally, a feature pyramid decoder aggregates the fused multi-scale features, gradually integrating high-level semantic information with low-level spatial details, ultimately outputting a preliminary fusion prediction map. This approach employs a top-down, stepwise fusion strategy, integrating cross-modal features from different levels. In summary, the calculation is as follows:

[0086]

[0087]

[0088]

[0089] The above formula concatenates two cross-modal features, RD3 and RD4 (upsampled by 2x), and then fuses them using a 3x3 convolution kernel to obtain the fused feature RD. 43Following this method, RD 43 RD2 is fused in the same way, and so on, to obtain aggregated features RD at different levels. 4321 Output the fused prediction map Sout.

[0090] Step 5: Multi-task collaborative training.

[0091] Design a composite loss function for end-to-end training, the loss function consisting of three parts:

[0092] First, the loss from graffiti supervision L sc That is, based on the initial graffiti annotations, the partial cross-entropy loss, local coherence loss, and structural consistency loss are calculated to ensure that the prediction results are consistent with the sparse annotations and maintain local smoothness.

[0093] The expression for the partial cross-entropy loss function is as follows:

[0094]

[0095] Where gt represents the set of pixels marked in the graffiti label. Let i be the true value of pixel i. Let be the predicted value for pixel i. This loss is used to supervise the prediction of the target map by using basic information from doodle annotations.

[0096] Due to the sparsity of graffiti annotation information, using only graffiti annotation information to supervise model training is unlikely to yield effective results. However, adjacent pixels often exhibit some correlation; local coherence loss can better preserve similar pixels in adjacent regions. The formula for calculating the local coherence loss is as follows:

[0097]

[0098] in Let i be the true value of pixel i. This indicates the position of pixel i. Indicates the color of pixel i. Represents the area around pixel i Adjacent areas within the range, express loss, To normalize the weights, and For hyperparameters, express Normalization.

[0099] Structural consistency loss enhances the consistency of target recognition across different input scales. By learning at different input scales, the model can improve its generalization ability. Its calculation formula is as follows:

[0100]

[0101] in This is the corresponding value after shrinking the size. These are the image values ​​corresponding to the downsampling. With a confidence threshold of 0.85, the SSIM loss focuses on local structural similarity. The final graffiti-supervised loss is obtained as follows:

[0102]

[0103] Second, single-modal prediction loss , That is, the cross-entropy loss, which is based on the initial pseudo-labels, and supervises the prediction maps output by the RGB and depth monomodal prediction modules.

[0104]

[0105]

[0106] in , For the visible light extended pseudo-label value and the depth extended pseudo-label value of pixel i, , Let i be the visible light prediction value and depth prediction value for pixel i.

[0107] Third, aggregated monitoring loss L ppa That is, based on the high-quality aggregated pseudo-labels generated in step 4, the weighted cross-entropy loss and weighted IoU loss are used to supervise the fusion prediction graph output by the prediction model.

[0108]

[0109] in For weighted cross-entropy loss, The weighted IoU loss is represented by R, where R is the predicted result. This is an aggregated pseudo-tag.

[0110] During the training phase, the composite loss is calculated:

[0111]

[0112] Where L sc To supervise the damage caused by graffiti, and For single-mode prediction loss, L ppa The aggregated supervised loss is calculated for Sout based on the aggregated pseudo-label Y. The prediction model is trained end-to-end using the Adam optimizer.

[0113] Step 6: Target detection.

[0114] The RGB-depth image pairs to be detected are input into the trained model, which directly outputs a segmentation mask of the camouflaged target. During testing, only the RGB-depth image pairs to be tested need to be input, and the trained model directly outputs a binary segmentation mask, thus completing the detection of the camouflaged target.

[0115] like Figure 5 The figure shows a comparison between the method of this invention (Our) and other fully supervised and weakly supervised methods. The method of this invention achieves overall performance that meets or even surpasses that of classic fully supervised methods at extremely low annotation costs. As can be seen from the figure, in extreme scenarios where the target and background are highly similar, compared to fully supervised methods (such as C2FNet and GT), which are prone to misclassifying a large number of environmental elements as targets, this method can more accurately define the target range and significantly reduce false detections. When facing scenarios with complex, elongated edges or multiple targets, this method, through multi-layer feature and multi-modal fusion, can better capture the overall framework and edge details of the target, while other methods (such as SINet and CRNet) are prone to losing details or incorrect fusion.

Claims

1. A weakly supervised camouflage target detection method based on graffiti annotation and depth information assistance, characterized in that, include: Step 1: Obtain an RGB image and a depth image containing the camouflaged target, and add graffiti annotations to the RGB image to obtain a graffiti-annotated image; Step 2: Process the RGB image and the depth image separately using the superpixel segmentation algorithm, and combine them with the doodle annotation image to generate initial RGB pseudo-labels and initial depth pseudo-labels; Step 3: Use the two original PVTv2 networks as single-modal prediction modules to process the RGB image and the depth image respectively, and generate predicted RGB pseudo-labels and predicted depth pseudo-labels; The predicted RGB pseudo-labels and predicted depth pseudo-labels are aggregated using a pixel adaptive thinning algorithm to obtain aggregated pseudo-labels; Step 4: Construct a prediction model, including an encoder for extracting image features, an aggregation module for fusing features, and a decoder for outputting a prediction map. Design the composite loss function of the prediction model and train the prediction model end-to-end based on a public dataset. The composite loss function includes: graffiti supervision loss based on the graffiti-annotated image, single-modal prediction loss based on the initial RGB pseudo-label and the initial depth pseudo-label, and aggregate supervision loss based on the aggregated pseudo-label; Step 5: Input the RGB image and depth image to be detected into the trained prediction model, and output the predicted image of the camouflaged target.

2. The weakly supervised camouflage target detection method according to claim 1, wherein processing the RGB image and the depth image using a superpixel segmentation algorithm comprises: The superpixel segmentation algorithm is applied to the RGB image to obtain several superpixel regions of the RGB image. The intersection of the superpixel regions of the graffiti annotation image and the RGB image is taken to obtain the initial RGB pseudo-label. A superpixel segmentation algorithm is applied to the depth image to obtain several superpixel regions of the depth image and the superpixel region of the depth image. The intersection of the superpixel regions of the graffiti annotation image and the depth image is taken to obtain the initial depth pseudo-label.

3. The weakly supervised camouflage target detection method according to claim 2, wherein when the single-modal prediction module processes the RGB image, it takes the RGB image as input, extracts several RGB image features layer by layer from low to high, uses the initial RGB pseudo-label obtained in step 2 as supervision, decodes the extracted highest-level RGB image features, and generates predicted RGB pseudo-labels; When the single-modal prediction module processes the depth image, it takes the depth image as input, extracts several depth image features layer by layer from low to high, uses the initial depth pseudo-label obtained in step 2 as supervision, decodes the extracted highest layer depth image features, and generates predicted depth pseudo-labels.

4. The weakly supervised camouflage target detection method according to claim 3, wherein the pixel adaptive thinning algorithm comprises: Using the predicted RGB pseudo-labels and the predicted depth pseudo-labels as input, and guided by the color and spatial affinity of the RGB image, iterative optimization and weighted averaging are performed to output aggregated pseudo-labels.

5. The weakly supervised camouflage target detection method according to claim 4, wherein the encoder of the prediction model is a PVTv2 network, extracting several layers of RGB image features from the RGB image input to the prediction model, and extracting several layers of depth image features from the depth image input to the prediction model, wherein the number of layers of the RGB image features is the same as the number of layers of the depth image features.

6. The weakly supervised camouflage target detection method according to claim 5, wherein the aggregation module of the prediction model includes a depthwise separable pyramid pooling module and several attention fusion modules, wherein the depthwise separable pyramid pooling module performs context enhancement on the highest-level features extracted by the encoder, and the attention fusion module fuses the RGB image features and depth image features of the same layer into a cross-modal fusion feature; the decoder of the prediction model aggregates the fusion features generated by the aggregation module from top to bottom and outputs a prediction map.

7. The weakly supervised camouflage target detection method according to claim 6, wherein the graffiti supervision loss includes: By using the prediction graph of the image-supervised prediction model with graffiti annotations, we obtain partial cross-entropy loss, local coherence loss, and structural consistency loss, which are then summed to obtain the graffiti supervision loss.

8. The weakly supervised camouflage target detection method according to claim 7, wherein the single-modal prediction loss includes: Based on the initial RGB pseudo-labels, the predicted RGB pseudo-labels output by the single-modal prediction module are supervised to obtain a cross-entropy loss; based on the initial depth pseudo-labels, the predicted depth pseudo-labels output by the single-modal prediction module are supervised to obtain another cross-entropy loss; the two cross-entropy losses are summed to obtain the graffiti supervision loss.

9. The weakly supervised camouflage target detection method according to claim 8, wherein the aggregated supervision loss includes: Based on the aggregated pseudo-labels described in step 4, the prediction graph of the prediction model is supervised to obtain the weighted cross-entropy loss and weighted IoU loss, which are then summed to obtain the aggregated supervision loss.

10. A weakly supervised camouflage target detection system based on graffiti annotation and depth information assistance, characterized in that, include: The data preparation module is used to acquire RGB images and depth images containing the camouflaged target, and to add graffiti annotations to the RGB images to obtain graffiti-annotated images; The initial pseudo-label module is used to process the RGB image and the depth image respectively using a superpixel segmentation algorithm, and combine the graffiti annotation image to generate initial RGB pseudo-labels and initial depth pseudo-labels; The aggregation pseudo-label module is used to treat the two original PVTv2 networks as single-modal prediction modules, process the RGB image and the depth image respectively, and generate predicted RGB pseudo-labels and predicted depth pseudo-labels. The predicted RGB pseudo-labels and predicted depth pseudo-labels are aggregated using a pixel adaptive thinning algorithm to obtain aggregated pseudo-labels; The prediction model includes an encoder for extracting image features, an aggregation module for fusing features, and a decoder for outputting a prediction map. The composite loss function of the prediction model is designed, and the prediction model is trained end-to-end based on a public dataset. The composite loss function includes: graffiti supervision loss based on the graffiti-annotated image, single-modal prediction loss based on the initial RGB pseudo-label and the initial depth pseudo-label, and aggregate supervision loss based on the aggregated pseudo-label; The output module is used to process the RGB image and depth image to be detected through the prediction model and output a predicted map of the camouflaged target.

Images