A camouflaged target detection method based on visual perception mechanism

Abstract

The present invention relates to a method for detecting camouflaged objects based on a visual perception mechanism. The method comprises the following steps: extracting features from multiple levels of a camouflaged object image; fusing the features from multiple levels to obtain a first-stage prediction result; performing feature extraction on the features extracted from the camouflaged object and the first-stage prediction result; sending the extracted features to a layered mixed ratio interaction module to obtain a second-stage prediction result; refining the second-stage prediction result using a self-attention refinement module; and further optimizing the features at the peer level using an X-shaped refinement module to achieve accurate detection results. This method addresses the problem of camouflaged object detection in computer vision and is effective and practical in various scenarios.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a method for detecting a camouflaged target in a complex background by using a detection system based on a visual perception mechanism. Background Art

[0002] Camouflaged object detection contributes to both the natural world and human society, and has significant research value. Camouflage objects can be broadly divided into two categories: natural camouflage and artificial camouflage. Natural camouflage objects refer to animals and plants that hide in the background environment to adapt to the environment and protect themselves. They have their own advantages (such as body shape and color), while artificial camouflage objects (such as military camouflage uniforms) occur in real life. Camouflage detection has many potential applications in medicine (such as polyp segmentation and lung infection segmentation), industry (such as surface defect detection of objects such as handicrafts and mechanical parts), agriculture (such as locust detection and tomato detection), military (such as camouflaged soldier detection), and art (such as landscape image detection). Therefore, research and development in camouflaged object detection is of great significance.

[0003] The goal of camouflaged object detection is to accurately and efficiently detect camouflaged targets that are "perfectly" embedded in the surrounding environment. Camouflaged object detection is more challenging than traditional object detection due to the high intrinsic similarity between targets and non-targets.

[0004] With the rapid development of deep learning, numerous solutions have been proposed for detecting camouflaged targets. Common strategies include background matching, mimicking the color and pattern of the environment, destructive coloring, and camouflaging body contours. By designing targeted network architectures, discriminative camouflaged object features can be effectively explored for camouflaged object detection. Methods that incorporate auxiliary tasks into joint learning and multi-task learning frameworks can extract valuable additional cues from shared features to significantly enhance the feature representation for camouflaged target detection. Although these methods have achieved promising results, they suffer from low recognition rates and poor model generalization performance in complex scenarios, which do not meet practical needs. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a method for detecting camouflaged targets based on a visual perception mechanism, which can also be applied to the detection of salient objects to achieve high-precision detection of targets. The technical solutions adopted by the present invention are:

[0006] 1. Extract multiple levels of features from the camouflaged object image. Using Swin-Transformer as the backbone network, we extract four levels of features (F1, F2, F3, and F4) from the camouflaged object image.

[0007] 2. Fuse features from multiple levels to obtain the prediction results of the first stage.

[0008] 2.1) Use one layer of convolution to reduce the number of channels of feature F4 and obtain the prediction result P4 of the fourth level in the first stage.

[0009] 2.2) Add feature F4 and feature F3 to obtain the first stage third level fusion feature and use one layer of convolution to obtain the first stage third level prediction result P3.

[0010] 2.3) The third-level fusion feature of the first stage is added to the feature F2 to obtain the second-level fusion feature of the first stage and a layer of convolution is used to obtain the second-level prediction result P1 of the first stage. The prediction result of the first stage is P1.

[0011] 3. Perform feature extraction on the features extracted from the camouflaged object and the prediction results of the first stage. Multiply the prediction result P1 of the first stage and the feature F2 to obtain the initial enhanced feature S1.

[0012] 4. The extracted features are sent to the hierarchical mixed ratio interaction module to obtain the prediction results of the second stage.

[0013] 4.1) Input the S1 feature into the hierarchical mixed ratio interaction module and divide it into several groups The groups interacted with each other in an iterative manner.

[0014] 4.2) After passing through a convolutional block, each group is divided into three feature sets, H1, H2, and H3. H1 is used to exchange information with the next group, while H2 and H3 are used for channel modulation. From another perspective, the iterative structure in the hierarchical mixed ratio interaction module can be equivalent to a core pyramid structure, resulting in the second-stage prediction result S2.

[0015] 5. Refine the results of the second stage prediction through the self-attention refinement module.

[0016] 5.1) First, feed S2 into a 3x3 convolutional layer to output the result S2′. Then, perform two parallel 3x3 convolutional layers. We multiply the output of one of the convolutions by S2′.

[0017] 5.2) The output is fused with the result of another convolutional unit through element-wise addition to obtain S3. Considering that not every pixel is representative in the channel dimension, the feature is maximized through the maximum self-attention function in the channel dimension, and the sub-maximized feature is added to S3 to output S4.

[0018] 6. The features of the peer layer are further optimized through the X-shaped refinement module to achieve accurate detection results. S4 is sent to the X-shaped refinement module, and the local information of the peer layer is combined with the generated camouflage map through X-shaped cross-over to output S5.

[0019] The present invention proposes a camouflaged target detection method based on a visual perception mechanism, which can effectively detect camouflaged targets, is applicable to the detection of various types of camouflaged targets, improves the recognition accuracy, and achieves excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solution proposed by the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art.

[0021] Figure 1 A system flow chart of the method proposed in the present invention;

[0022] Figure 2 This is a structural diagram of the hierarchical mixing ratio interaction module for camouflaged target detection of the present invention;

[0023] Figure 3 This is a structural diagram of the self-attention refinement module for camouflaged target detection of the present invention;

[0024] Figure 4 This is a structural diagram of the X-shaped refinement module for camouflaged target detection according to the present invention;

[0025] Figure 5 The following is a comparison chart of the results of an embodiment of the present invention (the first column is the original image, the second column is the true value image, and the third column is the result image of this method). DETAILED DESCRIPTION

[0026] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0027] In this embodiment, if Figure 1 As shown, an embodiment of the present invention provides a method for detecting camouflaged targets based on a visual perception mechanism, and the specific steps include:

[0028] 1. Extract multiple levels of features from the camouflaged object image. Using Swin-Transformer as the backbone network, we extract four levels of features (F1, F2, F3, and F4) from the camouflaged object image.

[0029] 2. Fuse features from multiple levels to obtain the prediction results of the first stage.

[0030] 2.1) Use one layer of convolution to reduce the number of channels of feature F4 and obtain the prediction result P4 of the fourth level in the first stage.

[0031] 2.2) Add feature F4 and feature F3 to obtain the first stage third level fusion feature and use one layer of convolution to obtain the first stage third level prediction result P3.

[0032] 2.3) The third-level fusion feature of the first stage is added to the feature F2 to obtain the second-level fusion feature of the first stage and a layer of convolution is used to obtain the second-level prediction result P1 of the first stage. The prediction result of the first stage is P1.

[0033] 3. A method for extracting features from the features extracted from the camouflaged object and the prediction results of the first stage, multiplying the prediction result P1 of the first stage and the feature F2 to obtain the initial enhanced feature S1.

[0034] 4. The extracted features are sent to the hierarchical mixed ratio interaction module to obtain the prediction results of the second stage.

[0035] 4.1) Input the S1 feature into the hierarchical mixing ratio interaction module, such as Figure 1 As shown, the number of channels is expanded by first passing through the 1x1 convolution layer, and then the S1 features are divided into several groups through the BN layer, Relu layer and partition layer. The groups interacted with each other in an iterative manner.

[0036] 4.2) After passing through a convolution block, each group is divided into three feature sets, namely H1, H2, and H3. H1 is used to exchange information with the next group, and H2 and H3 are used for channel modulation.

[0037] 4.3) A feature modulation vector is obtained through PCRCS (representing pooling layer, convolution layer, ReLU layer, convolution layer, and sigmoid function), which is used to weight another cascade of features. This iterative hybrid strategy strives to learn key cues from different channels and obtain a powerful feature representation. From another perspective, the iterative structure in the hierarchical hybrid proportional interaction module can be equivalent to a core pyramid structure, resulting in the second-stage prediction result S2.

[0038] 5. Refine the results of the second stage prediction through the self-attention refinement module.

[0039] 5.1) In this embodiment, if Figure 3 As shown, S2 is first fed into the 3x3 convolutional layer output bit S2′.

[0040] 5.2) Two 3x3 convolutional layers are performed in parallel. In order to capture more complete information to the greatest extent possible, we multiply the output of one of the convolutions by S2′ element-wise.

[0041] 5.3) The multiplication result is added element-by-element with another convolution. Considering that not every pixel is representative in the channel dimension, the maximized feature S3 is obtained through the maximum self-attention function in the channel dimension. Finally, the result of the element-by-element addition is added element-by-element with the maximized feature S3 to output S4.

[0042] 5.4)S5=mul(S2,Conv1(S2'))+Conv2(S2')+M-attention(S3)

[0043] 6. The features of the peer layer are further optimized through the X-shaped refinement module to achieve accurate detection results.

[0044] 6.1) In this embodiment, if Figure 4 As shown, first S4 and feature F i It passes through the convolution layer, BN layer and ReLU layer in sequence, and then enters the 3x3 convolution layer respectively, and two output results are obtained.

[0045] 6.2) Assume that the output of S4 after the convolution layer, BN layer and ReLU layer is S 41 , the output of the 3x3 convolutional layer is S 42 .

[0046] 6.3) The output of the F feature through the convolution layer, BN layer and Relu layer is F1, and the output of the 3x3 convolution layer is F2. At this time, X-shaped refinement is performed to output S 41 Multiply the output F2 element by element and output S 42 The output F1 is multiplied element-by-element, and finally the two results of the element-by-element multiplication are connected in series to obtain the result S5.

[0047] 6.4)S 41 =CBR(S4);F1=CBR(F);S 42 =Conv 3x3 (S 41 );

[0048] 6.5) F2 = Conv 3x3 (F1);

[0049] 7. Pass Figure 5 As can be seen from the result comparison chart, the first column is the original image, the second column is the true value image, and the third column is the result image of this method.

[0050] The technical solution provided in this embodiment can accurately complete camouflaged target detection in complex scenes and low recognition rates.

[0051] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting camouflaged targets based on a visual perception mechanism, comprising the following steps: Extracting multiple levels of features from camouflaged object images; Fuse features from multiple levels to obtain the prediction results of the first stage; Perform feature extraction on the features extracted from the camouflaged object and the prediction results of the first stage; The extracted features are sent to the hierarchical mixed ratio interaction module to obtain the prediction results of the second stage; Refine the results of the second-stage predictions through the self-attention refinement module; The features of the peer layer are further optimized through the X-shaped refinement module to achieve accurate detection results; The method for extracting multiple levels of features of the camouflaged object image is as follows: using Swin-Transformer as the backbone network to extract four levels of features F1, F2, F3 and F4 of the camouflaged object image; The step fuses features of multiple levels to obtain the prediction result of the first stage by: using a layer of convolution to reduce the number of channels of feature F4 to obtain the prediction result P4 of the fourth level of the first stage, adding feature F4 and feature F3 to obtain the third level fused feature of the first stage and using a layer of convolution to obtain the prediction result P3 of the third level of the first stage, adding the third level fused feature of the first stage and feature F2 to obtain the second level fused feature of the first stage and using a layer of convolution to obtain the prediction result P1 of the second level of the first stage, and the prediction result of the first stage is P1; The method of extracting features from the features extracted from the camouflaged object and the prediction results of the first stage is as follows: multiplying the prediction result P1 of the first stage and the feature F2 to obtain the initial enhanced feature S1; The step sends the extracted features to the hierarchical mixing ratio interaction module, and the method for obtaining the prediction results of the second stage is: input the S1 features into the hierarchical mixing ratio interaction module and divide them into several groups The groups interact with each other in an iterative manner. After passing through a convolution block, each group is divided into three feature sets, H1, H2, and H3. H1 is used to exchange information with the next group, and H2 and H3 are used for channel modulation. From another perspective, the iterative structure in the hierarchical mixing ratio interaction module can be equivalent to a core pyramid structure, thereby obtaining the prediction result S2 of the second stage; The method for refining the results of the second-stage prediction through the self-attention refinement module is as follows: S2 is first sent to the 3x3 convolution layer output result S2′, and then two parallel 3x3 convolution layers are performed at the same time. The output of one of the convolutions is multiplied by S2′; then, the output is fused with the result of the other convolution unit through element addition to obtain S3; considering that not every pixel is representative in the channel dimension, the feature is maximized through the maximum self-attention function in the channel dimension, and the sub-maximized feature is added to S3 to output S4.

2. The method for detecting camouflaged targets based on a visual perception mechanism according to claim 1, characterized in that: The step further optimizes the features of the peer layer through the X-shaped refinement module to achieve accurate detection results: S4 is sent to the X-shaped refinement module, and the local information of the peer layer is combined with the generated camouflage map through X-shaped crossover to output S5.

Images