Method for image camouflage target detection

Abstract

The application provides a method for image camouflage target detection, comprising the following steps: inputting an original image into a Resnet model to obtain a second feature map to a fourth feature map and a final feature map; inputting the original image into a Swin-Transformer model to obtain a ninth feature map; fusing the final feature map and the ninth feature map to obtain a fifth feature map; generating a boundary according to the second feature map to the fifth feature map; obtaining a second camouflage object prediction map according to the boundary and the second feature map to the fifth feature map, and the second camouflage object prediction map is a final camouflage object prediction map. By using the Resnet model and the Swin-Transformer model, the camouflage target detection task for the RGB image can be well completed, and the complete and contour-accurate camouflage target can be accurately detected.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to a method for detecting camouflaged targets in images. Background Technology

[0002] Camouflaged target detection aims to develop methods capable of detecting camouflaged targets, thereby enhancing the robustness and reliability of existing target detection technologies and expanding their application scope to better serve social and public safety. In image camouflaged target detection tasks, accurately and completely identifying object details is highly challenging due to the complexity of the camouflaged object's shape and the high similarity between its contour and the surrounding environment.

[0003] Existing methods for detecting camouflaged targets simply use convolutional networks such as ResNet and VGG when extracting initial features, which ignores some global semantic information. However, using Transformers and their variations, or simply adding Transformers after convolution and low-level feature extraction, focuses too much on modeling the global context. This results in low-resolution features lacking detailed localization information, leading to coarse segmentation results and affecting detection outcomes.

[0004] In view of this, the present invention proposes a method for detecting camouflaged targets in images. Summary of the Invention

[0005] To address the aforementioned technical problem of poor detection results for existing camouflaged targets, this invention provides a method for detecting camouflaged targets in images. This invention primarily achieves relatively accurate camouflaged target detection by utilizing ResNet models, Swin-Transformer models, channel attention, and spatial attention modules.

[0006] The technical means employed in this invention are as follows:

[0007] This invention provides a method for detecting camouflaged targets in images, comprising:

[0008] The original image is input into the ResNet model to obtain the first feature map, the second feature map, the third feature map, the fourth feature map, and the final feature map in sequence;

[0009] The original image is input into the Swin-Transformer model to obtain the 6th, 7th, 8th, and 9th feature maps;

[0010] The final feature map and the 9th feature map are fused to obtain the 5th feature map. Among the 1st to 4th feature maps and the final feature map, the absolute value of the size difference between the final feature map and the 9th feature map is the smallest.

[0011] A boundary is generated based on the second feature map, the third feature map, the fourth feature map, and the fifth feature map;

[0012] Based on the boundary, the second feature map, the third feature map, the fourth feature map, and the fifth feature map, a second camouflage object prediction map corresponding to the second feature map, a third camouflage object prediction map corresponding to the third feature map, and a fourth camouflage object prediction map corresponding to the fourth feature map are obtained.

[0013] The second camouflage object prediction map is the final camouflage object prediction map.

[0014] Further, the final feature map and the 9th feature map are fused to obtain the 5th feature map, including:

[0015] The final feature map is input into the first channel attention module to obtain the channel attention map and the channel refinement features;

[0016] The channel attention map, the channel refinement feature, and the 9th feature map are input into the 1st spatial attention module to obtain the 5th feature map.

[0017] Further, generating the boundary based on the second feature map, the third feature map, the fourth feature map, and the fifth feature map includes:

[0018] The second feature map, the third feature map, the fourth feature map, and the fifth feature map are respectively input into the feature enhancement module to obtain the corresponding second enhanced feature, third enhanced feature, fourth enhanced feature, and fifth enhanced feature;

[0019] The second enhancement feature is convolved to obtain the second transition feature; the third enhancement feature is convolved to obtain the third transition feature; the fourth enhancement feature is upsampled and then convolved to obtain the fourth transition feature; and the fifth enhancement feature is upsampled and then convolved to obtain the fifth transition feature.

[0020] The first combination is obtained by concatenating the second transition feature and the fifth transition feature and then performing two consecutive convolution operations; the second combination is obtained by concatenating the third transition feature and the fourth transition feature and then performing two consecutive convolution operations.

[0021] A first transition combination is obtained by performing a reverse attention mechanism based on the first combination, and a second transition combination is obtained by performing a reverse attention mechanism based on the second combination;

[0022] The first combination and the first transition combination are multiplied element by element to obtain the first sub-combination, the second combination and the second transition combination are multiplied element by element to obtain the second sub-combination, and the first sub-combination and the second sub-combination are added element by element and then the ReLU function is performed to obtain the boundary.

[0023] Further, based on the boundary, the second feature map, the third feature map, the fourth feature map, and the fifth feature map, a second camouflage object prediction map corresponding to the second feature map, a third camouflage object prediction map corresponding to the third feature map, and a fourth camouflage object prediction map corresponding to the fourth feature map are obtained, including:

[0024] Input the i-th feature map into the i-th channel attention module, where 2≤i≤5;

[0025] The i-th channel attention module acquires the boundary and the i-th feature map to generate the i-th channel map and the i-th channel feature;

[0026] The i' spatial attention module acquires the i-th channel image, the i-th channel feature, and the (i+1)-th channel image and performs calculations to obtain the i-th prediction result. The i-th prediction result is then convolved to obtain the i-th camouflage object prediction image, where i' = i and 2 ≤ i' ≤ 4.

[0027] Further, the i'-th spatial attention module acquires the i-th channel image, the i-th channel feature, and the (i+1)-th channel image and calculates the i-th prediction result, which is obtained in the following manner:

[0028]

[0029] Where, x i For the i-th prediction result, R′ i Let x be the feature of the i-th channel. i+1 For the (i+1)th prediction result, A i For the i-th channel image, A i+1 This is the (i+1)th channel image. For element-wise addition, sigmoid is the sigmoid function operation, SA i′ The attention module operation is performed for the i'th spatial region.

[0030] Furthermore, it also includes:

[0031] The total loss of the second, third, and fourth camouflaged object prediction maps is calculated for mask supervision, and obtained as follows:

[0032]

[0033] Where Ltotal is the total loss, L i Let E be the loss for predicting the i-th masquerading object, 2≤i≤4, E be the boundary, Ldice be the attention loss for the boundary, AM be the attention matrix, and L be the boundary value. AG G is the attention-guided loss for the attention matrix. e G0 is the ground truth map of the camouflaged target boundary.

[0034] Furthermore, the loss of the prediction graph of the i-th masquerading object is obtained in the following manner:

[0035] L i =LwBCE(P i ,G0)+LwIOU(P i ,G0);

[0036] Among them, L i P is the loss of the prediction graph of the i-th camouflaged object. i Let LwBCE be the predicted graph of the i-th masquerading object, LwIOU be the weighted binary cross-entropy loss, and LwIOU be the weighted IOU loss.

[0037] Furthermore, the attention matrix is ​​obtained as follows:

[0038] AM = ReLU(BN(Conv(F+F)) T )));

[0039] Where AM is the attention matrix, ReLU is the ReLU function operation, Conv is the convolution operation, and F is the fifth feature map. T The transpose of the fifth feature map is given, and BN is the batch normalization operation.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] 1. The method for image camouflage target detection provided by this invention can be well trained for camouflage target detection tasks for RGB images by using ResNet and Swin-Transformer models. It can accurately detect camouflage target markers with complete targets and accurate contours, and handle challenging detection tasks with complex structures and objects.

[0042] 2. The method for detecting camouflaged targets in images provided by this invention makes the generated boundaries more accurate and robust through texture enhancement and inverse attention mechanisms.

[0043] 3. The method for image camouflage target detection provided by the present invention introduces a channel attention module and a spatial attention module, which combine the boundary with the features of the camouflage object in each layer, captures the long-range dependence of channel and spatial location, and enhances the features from a global perspective to obtain the final camouflage object prediction map. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic flowchart of a method for detecting camouflaged targets in images provided by the present invention.

[0046] Figure 2 This is another flowchart illustrating the method for detecting camouflaged targets in images provided by the present invention.

[0047] Figure 3 This is a schematic diagram for generating the fifth feature map.

[0048] Figure 4 This is a schematic diagram of a feature enhancement module.

[0049] Figure 5 This is a qualitative analysis comparison diagram of the image camouflage target detection method provided by the present invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0052] Combination Figure 1 and Figure 2 , Figure 1 This is a schematic flowchart of a method for detecting camouflaged targets in images provided by the present invention. Figure 2 This is another flowchart illustrating a specific embodiment of the method for detecting camouflaged targets in images provided by the present invention, comprising:

[0053] The original image is input into the ResNet model to obtain the first feature map, the second feature map, the third feature map, the fourth feature map, and the final feature map in sequence;

[0054] The original image is input into the Swin-Transformer model to obtain the 6th, 7th, 8th and 9th feature maps;

[0055] The final feature map and the 9th feature map are fused to obtain the 5th feature map. Among the 1st to 4th feature maps and the final feature map, the absolute value of the size difference between the final feature map and the 9th feature map is the smallest.

[0056] Generate the boundary based on the second, third, fourth, and fifth feature maps;

[0057] Based on the boundary, the second feature map, the third feature map, the fourth feature map, and the fifth feature map, we obtain the second camouflaged object prediction map corresponding to the second feature map, the third camouflaged object prediction map corresponding to the third feature map, and the fourth camouflaged object prediction map corresponding to the fourth feature map;

[0058] The second camouflage object prediction map is the final camouflage object prediction map.

[0059] Understandably, the first, second, third, and fourth feature maps, as well as the final feature map, are generated by the Res2Net-50 encoder. Since the first feature map is too close to the original image and contains too much redundant information, its information is not used in this method. The size of the i-th feature map is:

[0060]

[0061] The final feature map size is:

[0062]

[0063] Where W is the width of the feature map, H is the height of the feature map, and C is the height of the feature map. i C1 represents the number of channels in the i-th feature map, and C5 represents the number of channels in the final feature map.

[0064] Since the embedding dimension of the Transformer model is fixed, the input size is downsampled to 224×224, resulting in a 9th feature map size of 14×14×512. The final feature map's feature size is close to that of the 9th feature map. The final feature map and the 9th feature map are then fused to obtain the fused feature map, which is the 5th feature map.

[0065] In some alternative embodiments, refer to Figure 2 and Figure 3 , Figure 3 As a schematic diagram for generating the fifth feature map, the fifth feature map is obtained by fusing the final feature map and the ninth feature map, including:

[0066] The final feature map is input into the first channel attention module to obtain the channel attention map and channel refinement features, as follows:

[0067]

[0068] in, This is an element-wise multiplication operation. For element-wise addition, R z The final feature map after processing is obtained by sequentially performing convolution, batch normalization (BN), and ReLU operations on the final feature map. The final feature map is obtained by average pooling. The final feature map is obtained through max pooling, where Sigmoid is the Sigmoid function operation, A is the channel attention map, and R is... r 1×k represents the channel refinement feature, and 1×k represents the size of the convolution kernel.

[0069] The channel attention map, channel refinement features, and the 9th feature map are input into the 1st spatial attention module to obtain the 5th feature map, which includes:

[0070] After downsampling the 9th feature map, it is concatenated and convolved with the channel refinement features to obtain R. The channels with high and low attention in R are sorted and divided into a first channel group and a second channel group in a 3:2 ratio. The reduced channels in the first and second channel groups are filled with 0 values ​​to ensure that the number of channels remains unchanged. The calculation is as follows:

[0071]

[0072] Where ConBNR represents the operations of sequential convolution, batch normalization, and ReLU functions, and R... im For the first channel group, R subim For the second channel group, Avg is the average pooling operation, Max is the global pooling operation, and F is the 5th feature map.

[0073] In some alternative embodiments, refer to Figure 2 and Figure 4 , Figure 4 This is a schematic diagram of a feature enhancement module, which generates boundaries based on the second, third, fourth, and fifth feature maps, including:

[0074] The second, third, fourth, and fifth feature maps are input into the feature enhancement module to obtain the corresponding second, third, fourth, and fifth enhanced features;

[0075] The second enhancement feature is convolved to obtain the second transition feature; the third enhancement feature is convolved to obtain the third transition feature; the fourth enhancement feature is upsampled and then convolved to obtain the fourth transition feature; and the fifth enhancement feature is upsampled and then convolved to obtain the fifth transition feature, in the following manner:

[0076]

[0077] Where, r i For the i-th enhanced feature, r′ i Let Up be the i-th transition feature, Up be the upsampling operation, and Conv3×3 be the convolution operation with a 3×3 convolution kernel.

[0078] The first combination is obtained by concatenating the second and fifth transition features and then performing two consecutive convolution operations. The second combination is obtained by concatenating the third and fourth transition features and then performing two consecutive convolution operations. The results are obtained as follows:

[0079]

[0080] Where, r 52 For the first combination, r 43 This is the second combination.

[0081] The first transition combination is obtained by applying a reverse attention mechanism based on the first combination, and the second transition combination is obtained by applying a reverse attention mechanism based on the second combination, as follows:

[0082]

[0083] Where, r′ 52 For the first transition combination, r′ 43 This is the second transitional combination.

[0084] Multiplying the first combination and the first transition combination element-wise yields the first sub-combination; multiplying the second combination and the second transition combination element-wise yields the second sub-combination; adding the first sub-combination and the second sub-combination element-wise and then performing a ReLU function operation yields the boundary, as follows:

[0085]

[0086] Here, E represents the boundary, and the generated boundary is made more accurate and robust through texture enhancement and inverse attention mechanisms.

[0087] Understandably, referring to Figure 4 The feature enhancement module includes four branches: the first branch includes a 1×1 convolutional kernel; the second branch includes cascaded 1×3 and 3×1 convolutional kernels; the third branch includes cascaded 1×5 and 5×1 convolutional kernels; and the fourth branch includes cascaded 1×7 and 7×1 convolutional kernels. The four parallel branches reduce the number of channels of the input i-th feature map to 32 before sending it to three parallel components, where 2≤i≤5. Each parallel component includes a 3×3 dilated convolution. During the dilation convolution, the first component adds 3 channels, the second component adds 5 channels, and the third component adds 7 channels. The input i-th feature map yields the corresponding i-th enhanced feature.

[0088] In some alternative embodiments, reference continues to be made to... Figure 1 and Figure 2 Based on the boundary, the second feature map, the third feature map, the fourth feature map, and the fifth feature map, a second camouflage object prediction map corresponding to the second feature map, a third camouflage object prediction map corresponding to the third feature map, and a fourth camouflage object prediction map corresponding to the fourth feature map are obtained, including:

[0089] Input the i-th feature map into the i-th channel attention module, where 2≤i≤5;

[0090] The attention module of the i-th channel obtains the boundary and the i-th feature map to generate the i-th channel map and the i-th channel feature;

[0091] The i' spatial attention module acquires the i-th channel image, the i-th channel feature, and the i+1-th channel image and performs calculations to obtain the i-th prediction result. The i-th prediction result is then convolved to obtain the i-th camouflaged object prediction image, where i' = i and 2 ≤ i' ≤ 4.

[0092] Understandably, the attention module for the i-th channel first performs element-wise multiplication on the boundary and the i-th feature map, and then adds additional skip connections to obtain the result. Obtained in the following way:

[0093]

[0094] The input to the i-th channel attention module is the i-th channel attention map, and the final feature map is the input to the 1-th channel attention module. The i-th channel attention module operates in the same way as the 1-th channel attention module. The i-th channel map is essentially the same as the channel attention map, and the i-th channel features are essentially the same as the channel refinement features. The i'-th spatial attention module operates in the same way as the 1-th spatial attention module. Through the channel attention module and the spatial attention module, the boundary is combined with the features of the camouflaged object in each layer to capture the long-range dependencies of channels and spatial locations, and the features are enhanced from a global perspective to obtain the final camouflaged object prediction map.

[0095] Specifically, the i'-th spatial attention module acquires the i-th channel image, the i-th channel feature, and the (i+1)-th channel image, and performs calculations to obtain the i-th prediction result, as follows:

[0096]

[0097] Where, x i For the i-th prediction result, R′ i For the i-th channel feature, x i+1 For the (i+1)th prediction result, A i For the i-th channel image, A i+1 This is the image for the (i+1)th channel. For element-wise addition, sigmoid is the sigmoid function operation, SA i′ This is the operation of the attention module in the i' space.

[0098] In some alternative embodiments, reference continues to be made to... Figure 1 and Figure 2 The method for detecting camouflaged targets in images provided by this invention further includes:

[0099] The total loss of the predicted maps of the second, third, and fourth camouflaged objects is calculated for mask supervision, and obtained as follows:

[0100]

[0101] Where Ltotal is the total loss, L i Let E be the loss for predicting the i-th disguised object, 2≤i≤4, E be the boundary value, Ldice be the boundary attention loss, AM be the attention matrix, and L be the boundary value. AG To achieve the attention-guided loss of the attention matrix, G e G0 is the ground truth map of the camouflaged target boundary.

[0102] Specifically, the loss of the prediction map for the i-th masquerading object is obtained as follows:

[0103] L i =LwBCE(P i ,G0)+LwIOU(P i ,G0);

[0104] Among them, L i P is the loss for predicting the i-th masquerading object. i Let LwBCE be the prediction graph of the i-th masquerading object, LwIOU be the weighted binary cross-entropy loss, and LwIOU be the weighted IOU loss.

[0105] Specifically, note the matrix, which is obtained as follows:

[0106] AM = ReLU(BN(Conv(F+F)) T )));

[0107] Where AM is the attention matrix, ReLU is the ReLU function operation, Conv is the convolution operation, F is the 5th feature map, and F T This is the transpose of the 5th feature map, and BN is the batch normalization operation.

[0108] In some alternative embodiments, refer to Tables 1 and 2, as follows:

[0109] Table 1. Quantitative Comparison of CAMO and CHAMELEON Datasets

[0110]

[0111] Table 2. Quantitative Comparison of COD10K and NC4K Datasets

[0112]

[0113] Each dataset uses four evaluation metrics, namely the structural similarity index (S... α ), mean E measure (E) φ) Weighted F-measure The mean absolute error (MAE) is shown in Tables 1 and 2. "↑" indicates that the larger the better, and "↓" indicates that the smaller the better. The best result in each group of indicators is shown in bold, and the second-ranked result is shown in italics. BANet (Ours) is the method for image camouflage target detection provided by this invention. As can be seen from Tables 1 and 2, the method provided by this invention can achieve relatively accurate camouflage target detection.

[0114] In some alternative embodiments, refer to Figure 5 , Figure 5 This is a qualitative analysis comparison of the image camouflage target detection method provided by the present invention. `image` represents the original image, and `Ours` represents the final camouflage object prediction image obtained using the image camouflage target detection method provided by the present invention. Figure 5 It can be seen that, compared with other methods, the method provided by the present invention can more clearly segment the disguised object.

[0115] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0116] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting camouflaged targets in images, characterized in that, include: The original image is input into the ResNet model to obtain the first feature map, the second feature map, the third feature map, the fourth feature map, and the final feature map in sequence; The original image is input into the Swin-Transformer model to obtain the 6th, 7th, 8th, and 9th feature maps; The final feature map and the 9th feature map are fused to obtain the 5th feature map. Among the 1st to 4th feature maps and the final feature map, the absolute value of the size difference between the final feature map and the 9th feature map is the smallest. The boundary is generated based on the second feature map, the third feature map, the fourth feature map, and the fifth feature map, including: The second feature map, the third feature map, the fourth feature map, and the fifth feature map are respectively input into the feature enhancement module to obtain the corresponding second enhanced feature, third enhanced feature, fourth enhanced feature, and fifth enhanced feature; The second enhancement feature is convolved to obtain the second transition feature; the third enhancement feature is convolved to obtain the third transition feature; the fourth enhancement feature is upsampled and then convolved to obtain the fourth transition feature; and the fifth enhancement feature is upsampled and then convolved to obtain the fifth transition feature. The first combination is obtained by concatenating the second transition feature and the fifth transition feature and then performing two consecutive convolution operations; the second combination is obtained by concatenating the third transition feature and the fourth transition feature and then performing two consecutive convolution operations. A first transition combination is obtained by performing a reverse attention mechanism based on the first combination, and a second transition combination is obtained by performing a reverse attention mechanism based on the second combination; The first combination and the first transition combination are multiplied element by element to obtain the first sub-combination, the second combination and the second transition combination are multiplied element by element to obtain the second sub-combination, the first sub-combination and the second sub-combination are added element by element and then the ReLU function is performed to obtain the boundary. Based on the boundary, the second feature map, the third feature map, the fourth feature map, and the fifth feature map, a second camouflage object prediction map corresponding to the second feature map, a third camouflage object prediction map corresponding to the third feature map, and a fourth camouflage object prediction map corresponding to the fourth feature map are obtained, including: Input the i-th feature map into the i-th channel attention module, where 2≤i≤5; The i-th channel attention module acquires the boundary and the i-th feature map to generate the i-th channel map and the i-th channel feature; The i'-th spatial attention module acquires the i-th channel image, the i-th channel feature, and the (i+1)-th channel image, and performs calculations to obtain the i-th prediction result. A convolution operation is then performed on the i-th prediction result to obtain the i-th camouflage object prediction image. =i and 2≤ ≤4; The second camouflage object prediction map is the final camouflage object prediction map.

2. The method for detecting camouflaged targets in images according to claim 1, characterized in that, The fifth feature map is obtained by fusing the final feature map and the ninth feature map, including: The final feature map is input into the first channel attention module to obtain the channel attention map and the channel refinement features; The channel attention map, the channel refinement feature, and the 9th feature map are input into the 1st spatial attention module to obtain the 5th feature map.

3. The method for detecting camouflaged targets in images according to claim 1, characterized in that, The i'-th spatial attention module acquires the i-th channel image, the i-th channel feature, and the (i+1)-th channel image, and calculates the i-th prediction result, which is obtained in the following manner: ； in, For the i-th prediction result, For the i-th channel feature, For the (i+1)th prediction result, This is the i-th channel image. This is the (i+1)th channel image. For element-wise addition, For Sigmoid function operations, The attention module operation is performed for the i'th spatial region.

4. The method for detecting camouflaged targets in images according to claim 1, characterized in that, Also includes: The total loss of the second, third, and fourth camouflaged object prediction maps is calculated for mask supervision, and obtained as follows: ； in, For the total loss, The loss is the prediction graph for the i-th masquerading object, where 2 ≤ i ≤ 4. For the boundary, The attention loss at the boundary, For attention matrix, The attention-guided loss for the attention matrix. To disguise the ground truth map of the target boundary, The ground truth graph for camouflaging targets.

5. The method for detecting camouflaged targets in an image according to claim 4, characterized in that, The loss of the prediction graph of the i-th masquerading object is obtained as follows: ； in, The loss of the predicted graph of the i-th camouflaged object. For the prediction map of the i-th camouflaged object, For weighted binary cross-entropy loss, This is the weighted IOU loss.

6. The method for detecting camouflaged targets in an image according to claim 5, characterized in that, The attention matrix is ​​obtained as follows: ； in, The attention matrix is... For ReLU function operations, For convolution operations, For the fifth feature map, This is the transpose of the fifth feature map. This is a batch normalization operation.

Images