Camouflage target detection method based on multi-channel coupling and context information

Abstract

The invention discloses a camouflage target detection method and device based on multichannel coupling and context information and a storage medium, and the method comprises the steps: obtaining a target image, determining the feature tensor of the target image according to a target network model, the feature tensor comprises a first feature tensor, a second feature tensor, a third feature tensor and a fourth feature tensor; inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflage target position; and based on the first camouflage target position, inputting the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor into a target locking module for iterative processing to obtain a second camouflage target position. Therefore, the detection precision of camouflage target detection is effectively improved, and a more accurate camouflage target position is obtained.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a method, device, electronic device and storage medium for detecting camouflaged targets based on multi-channel coupling and context information. Background Art

[0002] In computer vision, image information is often divided into shallow features and deep features. Deep features are also called high-level features. They have rich semantic information and can express image properties that are easier for observers to understand, but they cannot accurately estimate the location of the target. Low-level features are close to the input image and contain more edge contour information, which can distinguish and locate the target more accurately.

[0003] Traditional camouflaged target detection technology mainly relies on manual judgment and recognition, or uses hand-crafted low-level features to capture subtle differences in texture, intensity and color. However, this method often performs poorly in complex and changing environments and has low efficiency. In addition, with the development of deep learning technology, camouflaged target detection technology based on deep learning has gradually become mainstream. Deep learning methods use deep learning algorithms such as convolutional neural networks to automatically learn and extract discriminant features to achieve automatic recognition and detection of camouflaged targets. This method has higher accuracy and robustness and can adapt to complex and changing environments. However, the detection technology in the existing technology has low detection accuracy, and the detection results of the camouflaged target position are less accurate. Summary of the invention

[0004] The embodiment of the present application solves the problem of low reliability optimization efficiency in the prior art by providing a camouflaged target detection method based on multi-channel coupling and context information.

[0005] In order to achieve the above object, the technical solution of the embodiment of the present invention is:

[0006] In a first aspect, an embodiment of the present invention provides a camouflaged target detection method based on multi-channel coupling and context information, comprising:

[0007] Acquire a target image, and determine a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor;

[0008] Inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflaged target position;

[0009] Based on the first camouflaged target position, the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor are respectively input into a target locking module for iterative processing to obtain a second camouflaged target position.

[0010] In some possible implementations, inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first disguised target position includes:

[0011] Inputting the first feature tensor and the second feature tensor into a first target positioning submodule respectively to obtain a first target feature tensor and a second target feature tensor, wherein the target positioning module includes the first target positioning submodule;

[0012] Inputting the first target feature tensor and the second target feature tensor into a second target positioning submodule respectively to obtain a third target feature tensor and a fourth target feature tensor;

[0013] The third target feature tensor and the fourth target feature tensor are fused to obtain a fusion result, and the fusion result is input into a deformation convolution module to obtain the first camouflaged target position.

[0014] In some possible implementations, respectively inputting the first feature tensor and the second feature tensor into a first target positioning submodule to obtain a first target feature tensor and a second target feature tensor includes:

[0015] Performing deformation and transposition processing on the first feature tensor to obtain a first sub-feature tensor and a second sub-feature tensor;

[0016] Multiplying the first sub-feature tensor and the second sub-feature tensor to obtain a first target positioning sub-matrix;

[0017] Using a normalized exponential function, normalizing the first target positioning submatrix to obtain a first normalized target positioning submatrix;

[0018] Multiplying the first normalized target positioning submatrix and the first feature tensor to obtain a first feature mapping matrix, and performing feature tensor reconstruction processing on the first feature mapping matrix to obtain a first reconstruction matrix;

[0019] Performing adaptive scaling processing on the first reconstruction matrix to obtain a first adaptive feature tensor, and performing pixel-by-pixel summation processing on the first adaptive feature tensor and the first feature tensor to obtain a first target feature tensor;

[0020] The second feature tensor is input into the first target positioning submodule to obtain a second target feature tensor.

[0021] In some possible implementations, inputting the second feature tensor into the first target positioning submodule to obtain the second target feature tensor includes:

[0022] Performing deformation and transposition processing on the second feature tensor to obtain a third sub-feature tensor and a fourth sub-feature tensor;

[0023] Multiplying the third sub-feature tensor and the fourth sub-feature tensor to obtain a second target positioning sub-matrix;

[0024] Using a normalized exponential function, normalizing the second target positioning submatrix to obtain a second normalized target positioning submatrix;

[0025] Multiplying the second normalized target positioning submatrix and the second feature tensor to obtain a second feature mapping matrix, and performing feature tensor reconstruction processing on the second feature mapping matrix to obtain a second reconstructed matrix;

[0026] The second reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a second adaptive feature tensor, and the second adaptive feature tensor and the second feature tensor are summed pixel by pixel to obtain the second target feature tensor.

[0027] In some possible implementations, respectively inputting the first target feature tensor and the second target feature tensor into a second target positioning submodule to obtain a third target feature tensor and a fourth target feature tensor includes:

[0028] Performing feature coupling processing on the first target feature tensor and the second target feature tensor respectively to obtain a first coupling feature and a second coupling feature;

[0029] Performing deformation and transposition processing on the first coupling feature to obtain a fifth sub-feature tensor and a sixth sub-feature tensor;

[0030] Multiplying the fifth sub-feature tensor and the sixth sub-feature tensor to obtain a third target positioning sub-matrix;

[0031] Using a normalized exponential function to normalize the third target positioning submatrix to obtain a third normalized target positioning submatrix;

[0032] Multiplying the third normalized target positioning submatrix and the first coupling feature to obtain a third feature mapping matrix, and performing feature tensor reconstruction on the third feature mapping matrix to obtain a third reconstruction matrix;

[0033] Performing adaptive scaling processing on the third reconstruction matrix to obtain a third adaptive feature tensor, and performing pixel-by-pixel summing processing on the third adaptive feature tensor and the third feature tensor to obtain the third target feature tensor;

[0034] The second coupling feature is input into the second target positioning submodule to obtain a fourth target feature tensor.

[0035] In some possible implementations, inputting the second coupling feature into the second target positioning submodule to obtain a fourth target feature tensor includes:

[0036] Performing deformation and transposition processing on the second coupling feature to obtain a seventh sub-feature tensor and an eighth sub-feature tensor;

[0037] Multiplying the seventh sub-feature tensor and the eighth sub-feature tensor to obtain a fourth target positioning sub-matrix;

[0038] Using a normalized exponential function to normalize the fourth target positioning submatrix to obtain a fourth normalized target positioning submatrix;

[0039] multiplying the fourth normalized target positioning submatrix and the second coupling feature to obtain a fourth feature mapping matrix, and performing feature tensor reconstruction on the fourth feature mapping matrix to obtain a fourth reconstruction matrix;

[0040] The fourth reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a fourth adaptive feature tensor, and the fourth adaptive feature tensor and the fourth feature tensor are summed pixel by pixel to obtain the fourth target feature tensor.

[0041] In some possible implementations, based on the first disguised target position, respectively inputting the first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor into a target locking module for iterative processing to obtain a second disguised target position includes:

[0042] Performing upsampling processing on the first camouflaged target position to obtain a first sampling result;

[0043] Performing convolution processing on the first sampling result, the second feature tensor, and the first convolution layer to obtain a first convolution result;

[0044] Performing convolution processing on the second feature tensor and the second convolution layer to obtain a second convolution result;

[0045] Performing convolution processing on the second feature tensor and the third convolution layer, and multiplying the convolution processing result by the first adaptive learning coefficient to obtain a third convolution result;

[0046] Performing convolution processing on the second feature tensor and the fourth convolution layer, and multiplying the convolution processing result by the second adaptive learning coefficient to obtain a fourth convolution result;

[0047] Performing upsampling processing on the first feature tensor, and multiplying the sampling processing result by the third adaptive learning coefficient to obtain a second sampling result;

[0048] Adding the first convolution result, the second convolution result, the third convolution result, the fourth convolution result and the second sampling result to obtain a second sub-camouflaged target position;

[0049] The second feature tensor, the third feature tensor and the fourth feature tensor are sequentially input into the target locking module for iterative processing to obtain a second camouflaged target position.

[0050] In a second aspect, an embodiment of the present invention provides a camouflaged target detection device based on multi-channel coupling and context information, comprising:

[0051] A determination module, used to obtain a target image, and determine a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor;

[0052] An input module, used for inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflaged target position;

[0053] The iteration module is used to input the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor into the target locking module for iterative processing based on the first camouflaged target position to obtain a second camouflaged target position.

[0054] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the method provided in the first aspect of the present invention when executing the executable instructions or computer programs stored in the memory.

[0055] In a fourth aspect, the present invention provides a computer-readable storage medium storing executable instructions for causing a processor to execute the method provided in the first aspect of the present invention.

[0056] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages:

[0057] In the embodiment of the present invention, the comprehensive importance is combined with the Mustang algorithm, the comprehensive importance is used to analyze the change law of system reliability with the component state probability, a local search rule based on the comprehensive importance is constructed, the local search rule is added to the Mustang algorithm, the first reliability optimization result obtained by the Mustang algorithm is optimized, and the second reliability optimization result is obtained. In this way, the global search characteristics of the Mustang algorithm and the local search characteristics of the comprehensive importance are fully utilized to improve the efficiency of reliability optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention, the accompanying drawings required for use in the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0059] Figure 1 Schematic diagram of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information provided for the implementation of the present invention Figure 1 ;

[0060] Figure 2 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 2 ;

[0061] Figure 3 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 3 ;

[0062] Figure 4 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 4 ;

[0063] Figure 5 It is a structural schematic diagram of a first embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention;

[0064] Figure 6 Schematic diagram of the structure of a camouflaged target detection device based on multi-channel coupling and context information in an embodiment of the present invention;

[0065] Figure 7 The figure is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0067] In the relevant description of this embodiment, the terms "including, containing, having" and the like are open terms and are generally understood to include but not be limited to; the term "at least one" is generally understood to mean one or more, where "plurality" refers to two or more; the term "at least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items, for example, "at least one of a, b or c", or "at least one of a, b and c", can all represent: a, b, c, ab (i.e., a and b), ac, bc, or abc, where a, b, c can be single or multiple, respectively; the symbol "A / B" is used to describe the selection relationship of associated objects, generally indicating an "or" relationship before and after.

[0068] In the following description of the present embodiment, the terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0069] Those skilled in the art should understand that in the following description of the embodiments of the present application, the order of serial numbers does not mean the order of execution, some or all of the steps can be executed in parallel or sequentially, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0070] Those skilled in the art will appreciate that the numerical ranges in the embodiments of the present application are to be construed as also specifically disclosing each intermediate value between the upper and lower limits of the scope. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the range is also included in the present invention. The upper and lower limits of these smaller ranges may be independently included or excluded in the scope.

[0071] Unless otherwise specified, the technical / scientific terms used herein have the same meanings as those generally understood by those skilled in the art to which this application belongs. Although this application only describes preferred methods and materials, any methods and materials similar or equivalent to these may also be used in the implementation or testing of this application. All documents mentioned in this specification are incorporated by reference to disclose and describe methods and / or materials related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0072] In order to illustrate the technical solution of the present invention, specific embodiments are provided below for illustration.

[0073] First, the terms involved in this application are explained:

[0074] The Residual Network-50 (ResNet-50) model introduces the concept of "residual learning". By adding residual connections to the network, information can be directly transferred between network layers, thereby effectively solving the problems of gradient vanishing and network degradation in deep neural networks. The application scenarios of this model are: image classification, target detection, and image segmentation.

[0075] The Position-locked Network (PLNet) is a camouflaged target positioning and locking network. This network is mainly used to achieve accurate detection of camouflaged targets. The detection process mainly includes two stages: target search and locking.

[0076] In computer vision, image information is often divided into shallow features and deep features. Deep features are also called high-level features. They have rich semantic information and can express image properties that are easier for observers to understand, but they cannot accurately estimate the location of the target. Low-level features are close to the input image and contain more edge contour information, which can distinguish and locate the target more accurately.

[0077] Traditional camouflaged target detection technology mainly relies on manual judgment and recognition, or uses hand-crafted low-level features to capture subtle differences in texture, intensity and color. However, this method often performs poorly in complex and changing environments and has low efficiency. In addition, with the development of deep learning technology, camouflaged target detection technology based on deep learning has gradually become mainstream. Deep learning methods use deep learning algorithms such as convolutional neural networks to automatically learn and extract discriminant features to achieve automatic recognition and detection of camouflaged targets. This method has higher accuracy and robustness and can adapt to complex and changing environments. However, the detection technology in the existing technology has low detection accuracy, and the detection results of the camouflaged target position are less accurate.

[0078] Based on this, an embodiment of the present invention provides a disguised target detection method based on multi-channel coupling and context information, which is used to solve the problem of low accuracy of the detection result of the disguised target position in the prior art.

[0079] Figure 1 Schematic diagram of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information provided for the implementation of the present invention Figure 1 , see Figure 1 As shown, the above-mentioned camouflaged target detection method based on multi-channel coupling and context information may include:

[0080] S101, acquiring a target image, and determining a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor.

[0081] Among them, the target image is an image to be used for camouflaged target detection. The feature tensor of the target image can be obtained by inputting the target image into a target network model. The target network model can be, for example, ResNet-50. The first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor can be, for example, respectively: a deep feature tensor, a sub-deep feature tensor, a sub-shallow feature tensor, and a shallow feature tensor.

[0082] S102: Input the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflaged target position.

[0083] Among them, the target positioning module can be, for example: using the deep feature tensor and the second deep feature tensor output by ResNet-50 as the input of the target positioning module respectively, using the rich semantic information of the deep feature tensor to preliminarily locate the target position, and then enhancing the semantic representation of the result of the second deep feature tensor information through upsampling operation, further assisting in generating a first camouflaged target position module, which first camouflaged target position can be, for example, an initialized target position.

[0084] S103: Based on the first camouflaged target position, the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor are respectively input into a target locking module for iterative processing to obtain a second camouflaged target position.

[0085] It can be understood that the target positioning module and the target locking module can also be combined into a module corresponding to the camouflaged target detection system, which module can be a camouflaged target positioning locking network (Position-locked Network, PLNet) module based on multi-channel feature structure coupling.

[0086] The camouflaged target detection method based on multi-channel coupling and context information provided in this embodiment first extracts four feature tensors through a target network model, then inputs a deep feature tensor and a sub-deep feature tensor of the four feature tensors into a target locking module to determine a first camouflaged target position, and then based on the first camouflaged target position, the feature tensors are sequentially input into the target locking module for iterative processing, thereby obtaining a more accurate second camouflaged target position. The method processes the feature tensors in layers, thereby determining the edge features of the image on the basis of determining the initial potential camouflaged target position, thereby greatly improving the accuracy of camouflaged target detection.

[0087] Figure 2 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 2 , this embodiment is in Figure 1 Based on the embodiment, the process of obtaining the first target feature tensor and the second target feature tensor in the camouflaged target detection method based on multi-channel coupling and context information is described in detail, see Figure 2 As shown, the above-mentioned camouflaged target detection method based on multi-channel coupling and context information may include:

[0088] S201, acquiring a target image, and determining a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor.

[0089] Step S201 is similar to the above-mentioned step S101 and will not be described in detail again.

[0090] S202: Perform deformation and transposition processing on the first feature tensor to obtain a first sub-feature tensor and a second sub-feature tensor.

[0091] The first feature tensor may be, for example, a deep feature tensor, and the first feature tensor may be expressed as:

[0092]

[0093] Among them, Fr is the first feature tensor, C, H and W represent the number of channels, height and width of the first feature tensor respectively. Represents the feature tensor matrix.

[0094] The first sub-feature tensor can be expressed as:

[0095]

[0096] Among them, Frc represents the first sub-feature tensor after the first feature tensor is deformed, N is the number of pixels in each channel, and its size is H×W.

[0097] The second sub-feature tensor can be expressed as:

[0098]

[0099] Wherein, Frc' represents the second sub-feature tensor after the first feature tensor is deformed and transposed.

[0100] It can be understood that before this step performs deformation and transposition processing on the first feature tensor, for example, the first feature tensor can also be firstly input into the first target positioning submodule corresponding to the target positioning module, and the subsequent processing flow is performed in this module. The first target positioning submodule can be, for example, a channel attention subblock.

[0101] S203: Multiply the first sub-feature tensor and the second sub-feature tensor to obtain a first target positioning sub-matrix.

[0102] S204: Use a normalized exponential function to normalize the first target positioning sub-matrix to obtain a first normalized target positioning sub-matrix.

[0103] The normalized exponential function may be, for example, a softmax function.

[0104] The first normalized target positioning sub-matrix can be expressed as:

[0105]

[0106] Among them, Mc(i,j) represents the influence of the i-th channel on the j-th channel.

[0107] S205 , multiplying the first normalized target positioning submatrix and the first feature tensor to obtain a first feature mapping matrix, and performing feature tensor reconstruction processing on the first feature mapping matrix to obtain a first reconstructed matrix.

[0108] S206, performing adaptive scaling processing on the first reconstruction matrix to obtain a first adaptive feature tensor, and performing pixel-by-pixel summing processing on the first adaptive feature tensor and the first feature tensor to obtain a first target feature tensor.

[0109] The first target feature tensor can be expressed as:

[0110]

[0111] Among them, γ represents the adaptive learning scale factor, represents the first reconstruction matrix after the feature tensor is reconstructed, Mc represents the first target positioning submatrix, and Fr represents the first feature tensor.

[0112] S207: Perform deformation and transposition processing on the second feature tensor to obtain a third sub-feature tensor and a fourth sub-feature tensor.

[0113] The representation of the second feature tensor in step S207 is similar to the representation of the first feature tensor in step S201, and will not be described in detail here.

[0114] The second feature tensor may be, for example, a sub-deep feature tensor.

[0115] S208, multiplying the third sub-feature tensor and the fourth sub-feature tensor to obtain a second target positioning sub-matrix.

[0116] S209: Use a normalized exponential function to normalize the second target positioning sub-matrix to obtain a second normalized target positioning sub-matrix.

[0117] S210, multiplying the second normalized target positioning sub-matrix and the second feature tensor to obtain a second feature mapping matrix, and performing feature tensor reconstruction processing on the second feature mapping matrix to obtain a second reconstructed matrix.

[0118] S211, perform scale factor adaptive processing on the second reconstruction matrix to obtain a second adaptive feature tensor, and perform pixel-by-pixel summation processing on the second adaptive feature tensor and the second feature tensor to obtain the second target feature tensor.

[0119] The processing flow of obtaining the second target feature tensor in step S207 to step S211 is similar to the processing flow of obtaining the first target feature tensor corresponding to step S202 to step S206, and will not be repeated here.

[0120] It can be understood that the processing flow of obtaining the second target feature tensor in steps S207 to S211 or the processing flow of obtaining the first target feature tensor corresponds to steps S202 to S206, which is the processing content of the target positioning module. The processing module can also be, for example: a feature coupling target positioning module based on dual-channel fusion.

[0121] It is expandable that before the first feature tensor and the second feature tensor are deformed and transposed, the present embodiment may, for example, first perform reduction processing on the first feature tensor and the second feature tensor to remove some redundant information, and then perform deformation and transposition processing on the first feature tensor and the second feature tensor after the reduction processing, so as to facilitate subsequent processing of the first feature tensor and the second feature tensor and improve the accuracy of camouflaged target detection.

[0122] The camouflaged target detection method based on multi-channel coupling and context information provided in this embodiment is to perform deformation and transposition processing on the first feature tensor, and then perform multiplication processing on the first sub-feature tensor and the second sub-feature tensor after the deformation and transposition processing, normalization processing, feature tensor reconstruction processing, scale factor adaptive processing and pixel-by-pixel summation processing to obtain the first target feature tensor. The second target feature tensor has the same processing flow as the first target feature tensor. Through this series of processing, the accuracy of the subsequent coupling processing to obtain the first camouflaged target position is greatly improved.

[0123] Figure 3 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 3 , this embodiment is in Figure 2 Based on the embodiment, the process of obtaining the third target feature tensor, the fourth target feature tensor, and then obtaining the first camouflaged target position in the camouflaged target detection method based on multi-channel coupling and context information is described in detail, see Figure 3 As shown, the above-mentioned camouflaged target detection method based on multi-channel coupling and context information may include:

[0124] S301, performing feature coupling processing on the first target feature tensor and the second target feature tensor respectively to obtain a first coupling feature and a second coupling feature.

[0125] Among them, feature coupling processing can be implemented by feature coupling components (FCC), which can be a semi-symmetric structure. The feature coupling component divides the input features into three groups of paths, where the first two groups of paths pass through 1×1 convolution layers and 3×3 convolution layers respectively. In order to expand the feature acceptance range, each group of paths also uses dilated convolution to increase the coverage area of ​​the convolution kernel, and the expansion rate of the dilated convolution is 3; the third group of paths only uses 1×1 convolution layers for calculation. Finally, in order to remove all negative values ​​of feature pixels and increase the learning performance of the network structure, all convolutions of the feature coupling component are followed by an activation function nonlinear layer, and the activation function can be, for example, ReLU (Rectified Linear Unit).

[0126] S302: Perform deformation and transposition processing on the first coupling feature to obtain a fifth sub-feature tensor and a sixth sub-feature tensor.

[0127] It can be understood that before this step deforms and transposes the first coupling feature, for example, the first coupling feature can also be first input into the second target positioning submodule corresponding to the target positioning module, and subsequent processing steps are performed in this module. The second target positioning submodule can be, for example, a spatial attention subblock.

[0128] S303: Multiply the fifth sub-feature tensor and the sixth sub-feature tensor to obtain a third target positioning sub-matrix.

[0129] S304: Use a normalized exponential function to normalize the third target positioning sub-matrix to obtain a third normalized target positioning sub-matrix.

[0130] The normalized exponential function may be, for example, a softmax function.

[0131] S305, multiplying the third normalized target positioning sub-matrix and the first coupling feature to obtain a third feature mapping matrix, and performing feature tensor reconstruction processing on the third feature mapping matrix to obtain a third reconstruction matrix.

[0132] S306, performing scale factor adaptive processing on the third reconstruction matrix to obtain a third adaptive feature tensor, and performing pixel-by-pixel sum processing on the third adaptive feature tensor and the third feature tensor to obtain the third target feature tensor.

[0133] S307, performing deformation and transposition processing on the second coupling feature to obtain a seventh sub-feature tensor and an eighth sub-feature tensor.

[0134] S308 , multiplying the seventh sub-feature tensor and the eighth sub-feature tensor to obtain a fourth target positioning sub-matrix.

[0135] S309: using a normalized exponential function to normalize the fourth target positioning sub-matrix to obtain a fourth normalized target positioning sub-matrix.

[0136] S310, multiplying the fourth normalized target positioning sub-matrix and the second coupling feature to obtain a fourth feature mapping matrix, and performing feature tensor reconstruction processing on the fourth feature mapping matrix to obtain a fourth reconstruction matrix.

[0137] S311, perform scale factor adaptive processing on the fourth reconstruction matrix to obtain a fourth adaptive feature tensor, and perform pixel-by-pixel sum processing on the fourth adaptive feature tensor and the fourth feature tensor to obtain the fourth target feature tensor.

[0138] It can be understood that the processing flow of obtaining the third target feature tensor and the fourth target feature tensor in steps S302 to S311 is similar to the processing flow of obtaining the first target feature tensor corresponding to steps S202 to S206, and will not be repeated here.

[0139] S312: Fusing the third target feature tensor and the fourth target feature tensor to obtain a fusion result, and inputting the fusion result into a deformation convolution module to obtain the first camouflaged target position.

[0140] The first disguised target position may be expressed as:

[0141]

[0142] Among them, def<·> indicates that the feature is subjected to a deformed convolution operation, Fcfs indicates the feature tensor obtained by the deformed transpose process, Fcf indicates the output of the feature coupling component, and Ms indicates the fourth target positioning submatrix.

[0143] It can be understood that the processing flow of obtaining the third target feature tensor in steps S302 to S306 or the processing flow of obtaining the fourth target feature tensor corresponding to steps S307 to S311 is also the processing content of the target positioning module, and the processing module can also be, for example: a feature coupling target positioning module based on dual-channel fusion.

[0144] It can be understood that since the feature tensor corresponding to the third target feature tensor is a sub-deep feature tensor, before the third target feature tensor and the fourth target feature tensor are fused, for example, the third target feature tensor can also be subjected to maximum pooling processing, so that the third target feature tensor and the fourth target feature tensor can be better integrated.

[0145] The camouflaged target detection method based on multi-channel coupling and context information provided in this embodiment couples the first target feature tensor and the second target feature tensor to obtain the first coupling feature and the second coupling feature, then processes the first coupling feature accordingly to obtain the third target feature tensor, and then processes the second coupling feature accordingly to obtain the fourth target feature tensor, so as to facilitate fusion processing based on the third target feature tensor and the fourth target feature tensor, and then perform subsequent deformation convolution processing, so as to obtain a more accurate first camouflaged target position, and by adding the coupling processing method, the negative values ​​in the feature pixels are eliminated, and the fourth target feature tensor is increased. The learning performance of the network structure is improved, and the method of coupling the convolution in the corresponding components and then connecting to the nonlinear layer makes the output features of the first target positioning submodule, namely the channel attention subblock, not only integrate the interdependence characteristics between channels, but also have the ability to extract high-level semantic features. While enriching the feature semantic representation, it improves the recognizability of the output features, thereby greatly improving the accuracy of camouflaged target detection; at the same time, the fusion results are input into the deformation convolution module, so that feature learning can effectively cope with the geometric deformation in the target and scene, which not only makes the training have rotational deformation characteristics, but also further enhances the ability of the receptive field and improves the accuracy of camouflaged target detection.

[0146] Figure 4 FIG. 1 is a flow chart of an embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention. Figure 4 , this embodiment is in Figure 3 Based on the embodiment, the process of obtaining the second camouflaged target position in the camouflaged target detection method based on multi-channel coupling and context information is described in detail, see Figure 4 As shown, the above-mentioned camouflaged target detection method based on multi-channel coupling and context information may include:

[0147] S401, performing upsampling processing on a first disguised target position to obtain a first sampling result.

[0148] S402, performing convolution processing on the first sampling result, the second feature tensor and the first convolution layer to obtain a first convolution result.

[0149] S403: Perform convolution processing on the second feature tensor and the second convolution layer to obtain a second convolution result.

[0150] S404: Perform convolution processing on the second feature tensor and the third convolution layer, and multiply the convolution processing result with the first adaptive learning coefficient to obtain a third convolution result.

[0151] S405: Perform convolution processing on the second feature tensor and the fourth convolution layer, and multiply the convolution processing result with the second adaptive learning coefficient to obtain a fourth convolution result.

[0152] Specifically, the structure of each of the first convolutional layer, the second convolutional layer, the third convolutional layer and the fourth convolutional layer may include, for example: a (2k-1)×(2k-1) convolutional layer, a convolutional layer with a dilation rate of (2k-1) and a convolution kernel size of 3×3, and a batch normalization and activation function.

[0153] S406: Up-sample the first feature tensor, and multiply the sampling result by the third adaptive learning coefficient to obtain a second sampling result.

[0154] S407: Add the first convolution result, the second convolution result, the third convolution result, the fourth convolution result, and the second sampling result to obtain a second sub-disguised target position.

[0155] S408: sequentially input the second feature tensor, the third feature tensor, and the fourth feature tensor into the target locking module for iterative processing to obtain a second camouflaged target position.

[0156] Specifically, the target locking module combines four feature layers, namely, the deep feature tensor, the sub-deep feature tensor, the sub-shallow feature tensor, and the shallow feature tensor, and iteratively refines the first camouflaged target position layer by layer. The iterative processing process may include, for example: the first layer iteration is: first input the sub-deep feature tensor to obtain the corresponding first sampling result, the first convolution result, the second convolution result, the third convolution result, and the fourth convolution result, and then input the deep feature tensor to obtain the corresponding second sampling result, and add and combine the first sampling result, the first convolution result, the second convolution result, the third convolution result, the fourth convolution result, and the second sampling result to obtain the first layer second camouflaged target position;

[0157] The second layer iteration is as follows: firstly, the second shallow layer feature tensor is input to obtain the corresponding first sampling result, first convolution result, second convolution result, third convolution result, and fourth convolution result, and then the second deep layer feature tensor is input to obtain the corresponding second sampling result, and the first sampling result, first convolution result, second convolution result, third convolution result, fourth convolution result, and second sampling result are added and combined to obtain the second camouflaged target position of the second layer;

[0158] The third layer iteration is as follows: first input the shallow feature tensor to obtain the corresponding first sampling result, first convolution result, second convolution result, third convolution result, and fourth convolution result, then input the second shallow feature tensor to obtain the corresponding second sampling result, add and combine the first sampling result, first convolution result, second convolution result, third convolution result, fourth convolution result and second sampling result to obtain the second camouflaged target position of the third layer, complete the iterative processing of the second camouflaged target position, and the second camouflaged target position of the third layer is the final second camouflaged target position.

[0159] Specifically, the first convolution result, the second convolution result, the third convolution result, and the fourth convolution result can be expressed as:

[0160]

[0161] Among them, X1 is the first convolution result, Up(PreR) represents the first sampling result, and Fs'r n is the feature tensor of the current iteration input, Lm1 represents the first convolution layer; X2 is the second convolution result, Lm2 represents the second convolution layer; X3 is the third convolution result, Lm3 represents the third convolution layer, λ 1 represents the first adaptive learning coefficient; X4 is the fourth convolution result, Lm4 represents the fourth convolution layer, λ 2 represents the second adaptive learning coefficient;

[0162] The second disguised target position can be expressed as:

[0163]

[0164] Among them, Up(PreR) means sampling PreR, λ 3 represents the third adaptive learning coefficient, and X is the result of adding the first convolution result, the second convolution result, the third convolution result, and the fourth convolution result.

[0165] It can be understood that the processing flow of obtaining the second disguised target position in step S401 to step S408 is the processing content of the target locking module, and the processing module can also be, for example, a target locking module based on information iteration.

[0166] It is expandable that before iteratively processing the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor, the present embodiment may, for example, first perform reduction processing on the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor to remove some redundant information, and perform deformation and transposition processing on the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor after the reduction processing, so as to facilitate the subsequent processing flow and improve the accuracy of camouflaged target detection.

[0167] Expandable, the verification of the accuracy of the camouflaged target detection results based on multi-channel coupling and context information in this embodiment can use the currently public network data set to train and test the network, and compare it with algorithms such as camouflage detection network algorithms, positioning and focusing network algorithms. The network data set may include, for example: a chameleon algorithm data set, a camouflaged object data set, etc. The specific verification process may include, for example: using multiple algorithms to conduct comparative experiments to determine whether the binary segmentation result obtained in this embodiment is closer to the true value image. Another verification of this embodiment can be verified, for example, through a network ablation experiment, by randomly selecting multiple images in the training set according to different target sizes and camouflage degrees, and visually analyzing the output features through ablation experiments. It can be seen from this verification experiment that after coupling processing and deformed convolution structure processing, a purer feature representation is obtained in the reinforcement learning of channel attention features and spatial attention features, which provides strong support for the target positioning module to realize camouflaged target positioning, making the feature difference between the camouflaged target and the background more obvious, and the edge contour information of the target is also more richly expressed.

[0168] The camouflaged target detection method based on multi-channel coupling and context information provided in this embodiment is to input the second feature tensor, the third feature tensor and the fourth feature tensor into the target locking module for iterative processing in sequence, so that the second camouflaged target position obtained after layer-by-layer iteration is more accurate. At the same time, multiple convolutional layers and a batch normalization and activation function are combined in each convolutional layer. The receptive field can be increased through expansion rates of different scales, and the four branches are spliced ​​and the global feature information is aggregated. Finally, the prediction of the camouflaged target is completed by combining deeper level features, so that the camouflaged target detection result is gradually accurate. The input feature layer of each target locking module gradually contains more edge contour information from deep to shallow. Therefore, iterative processing to obtain the final camouflaged target detection result is also dynamically changing, and as the feature tensor is shallower, the difference between the target and the background is more obvious, and the camouflaged target detection result is more accurate.

[0169] Figure 5 FIG. 1 is a structural diagram of a first embodiment of a camouflaged target detection method based on multi-channel coupling and context information in an embodiment of the present invention, see Figure 5 shown.

[0170] Among them, Input Image represents the input target image, layer1, layer2, layer3 and layer4 are shallow feature tensors, sub-shallow feature tensors, sub-deep feature tensors and deep feature tensors respectively, PM represents the target positioning module, PreR represents the first camouflaged target position, LM represents the target locking module, and Ouput represents the second camouflaged target position after iterative processing. Specifically, the sub-deep feature tensor passes through the target positioning module and then enters the Maxpool module for maximum pooling processing, and the maximum pooling processing result is combined with the output of the deep feature tensor after the target positioning module to form the first camouflaged target position PreR, and then the first camouflaged target position is input into the LM to represent the target locking module for iterative processing to obtain the second camouflaged target position.

[0171] Based on the same inventive concept, an embodiment of the present application also provides a camouflaged target detection device based on multi-channel coupling and context information. Figure 6 FIG. 1 is a schematic diagram of a camouflaged target detection device based on multi-channel coupling and context information in an embodiment of the present invention. Figure 6 As shown, the camouflaged target detection device 600 based on multi-channel coupling and context information may include: a determination module 601, used to obtain a target image, and determine a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor;

[0172] An input module 602, configured to input the first feature tensor and the second feature tensor into a target positioning module to obtain a first disguised target position;

[0173] The iteration module 603 is used to input the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor into the target locking module for iterative processing based on the first disguised target position to obtain a second disguised target position.

[0174] In some possible implementations, the input module 602 is further used to input the first feature tensor and the second feature tensor into a first target positioning submodule respectively to obtain a first target feature tensor and a second target feature tensor, and the target positioning module includes the first target positioning submodule;

[0175] Inputting the first target feature tensor and the second target feature tensor into a second target positioning submodule respectively to obtain a third target feature tensor and a fourth target feature tensor;

[0176] The third target feature tensor and the fourth target feature tensor are fused to obtain a fusion result, and the fusion result is input into a deformation convolution module to obtain the first camouflaged target position.

[0177] In some possible implementations, the apparatus further includes: a processing module 604;

[0178] The processing module 604 is used to perform deformation and transposition processing on the first feature tensor to obtain a first sub-feature tensor and a second sub-feature tensor;

[0179] Multiplying the first sub-feature tensor and the second sub-feature tensor to obtain a first target positioning sub-matrix;

[0180] Using a normalized exponential function, normalizing the first target positioning submatrix to obtain a first normalized target positioning submatrix;

[0181] Multiplying the first normalized target positioning submatrix and the first feature tensor to obtain a first feature mapping matrix, and performing feature tensor reconstruction processing on the first feature mapping matrix to obtain a first reconstruction matrix;

[0182] Performing adaptive scaling processing on the first reconstruction matrix to obtain a first adaptive feature tensor, and performing pixel-by-pixel summation processing on the first adaptive feature tensor and the first feature tensor to obtain a first target feature tensor;

[0183] The second feature tensor is input into the first target positioning submodule to obtain a second target feature tensor.

[0184] In some possible implementations, the processing module 604 is further configured to perform a deformation transposition process on the second feature tensor to obtain a third sub-feature tensor and a fourth sub-feature tensor;

[0185] Multiplying the third sub-feature tensor and the fourth sub-feature tensor to obtain a second target positioning sub-matrix;

[0186] Using a normalized exponential function, normalizing the second target positioning submatrix to obtain a second normalized target positioning submatrix;

[0187] Multiplying the second normalized target positioning submatrix and the second feature tensor to obtain a second feature mapping matrix, and performing feature tensor reconstruction processing on the second feature mapping matrix to obtain a second reconstructed matrix;

[0188] The second reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a second adaptive feature tensor, and the second adaptive feature tensor and the second feature tensor are summed pixel by pixel to obtain the second target feature tensor.

[0189] In some possible implementations, the processing module 604 is further configured to perform feature coupling processing on the first target feature tensor and the second target feature tensor respectively to obtain a first coupling feature and a second coupling feature;

[0190] Performing deformation and transposition processing on the first coupling feature to obtain a fifth sub-feature tensor and a sixth sub-feature tensor;

[0191] Multiplying the fifth sub-feature tensor and the sixth sub-feature tensor to obtain a third target positioning sub-matrix;

[0192] Using a normalized exponential function to normalize the third target positioning submatrix to obtain a third normalized target positioning submatrix;

[0193] Multiplying the third normalized target positioning submatrix and the first coupling feature to obtain a third feature mapping matrix, and performing feature tensor reconstruction on the third feature mapping matrix to obtain a third reconstruction matrix;

[0194] Performing adaptive scaling processing on the third reconstruction matrix to obtain a third adaptive feature tensor, and performing pixel-by-pixel summing processing on the third adaptive feature tensor and the third feature tensor to obtain the third target feature tensor;

[0195] The second coupling feature is input into the second target positioning submodule to obtain a fourth target feature tensor.

[0196] In some possible implementations, the processing module 604 is further configured to perform a deformation transposition process on the second coupling feature to obtain a seventh sub-feature tensor and an eighth sub-feature tensor;

[0197] Multiplying the seventh sub-feature tensor and the eighth sub-feature tensor to obtain a fourth target positioning sub-matrix;

[0198] Using a normalized exponential function to normalize the fourth target positioning submatrix to obtain a fourth normalized target positioning submatrix;

[0199] multiplying the fourth normalized target positioning submatrix and the second coupling feature to obtain a fourth feature mapping matrix, and performing feature tensor reconstruction on the fourth feature mapping matrix to obtain a fourth reconstruction matrix;

[0200] The fourth reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a fourth adaptive feature tensor, and the fourth adaptive feature tensor and the fourth feature tensor are summed pixel by pixel to obtain the fourth target feature tensor.

[0201] In some possible implementations, the iteration module 603 is further configured to perform upsampling processing on the first disguised target position to obtain a first sampling result;

[0202] Performing convolution processing on the first sampling result, the second feature tensor, and the first convolution layer to obtain a first convolution result;

[0203] Performing convolution processing on the second feature tensor and the second convolution layer to obtain a second convolution result;

[0204] Performing convolution processing on the second feature tensor and the third convolution layer, and multiplying the convolution processing result by the first adaptive learning coefficient to obtain a third convolution result;

[0205] Performing convolution processing on the second feature tensor and the fourth convolution layer, and multiplying the convolution processing result by the second adaptive learning coefficient to obtain a fourth convolution result;

[0206] Performing upsampling processing on the first feature tensor, and multiplying the sampling processing result by the third adaptive learning coefficient to obtain a second sampling result;

[0207] Adding the first convolution result, the second convolution result, the third convolution result, the fourth convolution result and the second sampling result to obtain a second sub-camouflaged target position;

[0208] The second feature tensor, the third feature tensor and the fourth feature tensor are sequentially input into the target locking module for iterative processing to obtain a second camouflaged target position.

[0209] Based on the same inventive concept, an embodiment of the present application provides an electronic device, which may be consistent with the camouflaged target detection method based on multi-channel coupling and context information in one or more of the above embodiments. Figure 7 FIG. 1 is a schematic diagram of the structure of an electronic device in an embodiment of the present invention, see Figure 7 As shown, the electronic device 700 may adopt general computer hardware, including a processor 701 and a memory 702 .

[0210] In some possible implementations, at least one processor may constitute any physical device having a circuit that performs a logical operation on one or more inputs. For example, at least one processor may include one or more integrated circuits, including application specific integrated circuits, microchips, microcontrollers, microprocessors, all or part of a central processing unit, a graphics processing unit, a digital signal processor, a field programmable gate array, or other circuits suitable for executing instructions or performing logical operations. The instructions executed by at least one processor may be preloaded, for example, into a memory integrated with or embedded in a controller, or may be stored in a separate memory. The memory may include a random access memory, a read-only memory, a hard disk, an optical disk, a magnetic medium, a flash memory, other permanent, fixed or volatile memory, or any other mechanism capable of storing instructions. In some embodiments, at least one processor may include more than one processor. Each processor may have a similar structure, or the processors may have different configurations that are electrically connected or disconnected from each other. For example, the processor may be a separate circuit or integrated in a single circuit. When more than one processor is used, the processor may be configured to operate independently or collaboratively. The processors may be coupled electrically, magnetically, optically, acoustically, mechanically, or by other means that allow them to interact.

[0211] Based on the same inventive concept, the present application provides a computer storage medium, which stores computer executable instructions. After the computer executable instructions are executed by a processor, a camouflaged target detection method based on multi-channel coupling and context information as described in one or more of the above embodiments can be implemented.

[0212] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from other embodiments.

[0213] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present application.

Claims

1. A camouflaged target detection method based on multi-channel coupling and context information, characterized in that: include: Acquire a target image, and determine a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor; Inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflaged target position; Based on the first camouflaged target position, the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor are respectively input into a target locking module for iterative processing to obtain a second camouflaged target position.

2. The method according to claim 1, characterized in that The step of inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first disguised target position includes: Inputting the first feature tensor and the second feature tensor into a first target positioning submodule respectively to obtain a first target feature tensor and a second target feature tensor, wherein the target positioning module includes the first target positioning submodule; Inputting the first target feature tensor and the second target feature tensor into a second target positioning submodule respectively to obtain a third target feature tensor and a fourth target feature tensor, wherein the target positioning module includes the second target positioning submodule; The third target feature tensor and the fourth target feature tensor are fused to obtain a fusion result, and the fusion result is input into a deformation convolution module to obtain the first camouflaged target position.

3. The method according to claim 2, characterized in that The step of inputting the first feature tensor and the second feature tensor into a first target positioning submodule to obtain a first target feature tensor and a second target feature tensor comprises: Performing deformation and transposition processing on the first feature tensor to obtain a first sub-feature tensor and a second sub-feature tensor; Multiplying the first sub-feature tensor and the second sub-feature tensor to obtain a first target positioning sub-matrix; Using a normalized exponential function, normalizing the first target positioning submatrix to obtain a first normalized target positioning submatrix; Multiplying the first normalized target positioning submatrix and the first feature tensor to obtain a first feature mapping matrix, and performing feature tensor reconstruction processing on the first feature mapping matrix to obtain a first reconstruction matrix; Performing adaptive scaling processing on the first reconstruction matrix to obtain a first adaptive feature tensor, and performing pixel-by-pixel summation processing on the first adaptive feature tensor and the first feature tensor to obtain a first target feature tensor; The second feature tensor is input into the first target positioning submodule to obtain a second target feature tensor.

4. The method according to claim 3, characterized in that The step of inputting the second feature tensor into the first target positioning submodule to obtain a second target feature tensor includes: Performing deformation and transposition processing on the second feature tensor to obtain a third sub-feature tensor and a fourth sub-feature tensor; Multiplying the third sub-feature tensor and the fourth sub-feature tensor to obtain a second target positioning sub-matrix; Using a normalized exponential function, normalizing the second target positioning submatrix to obtain a second normalized target positioning submatrix; Multiplying the second normalized target positioning submatrix and the second feature tensor to obtain a second feature mapping matrix, and performing feature tensor reconstruction processing on the second feature mapping matrix to obtain a second reconstructed matrix; The second reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a second adaptive feature tensor, and the second adaptive feature tensor and the second feature tensor are summed pixel by pixel to obtain the second target feature tensor.

5. The method according to claim 2, characterized in that: The step of inputting the first target feature tensor and the second target feature tensor into a second target positioning submodule to obtain a third target feature tensor and a fourth target feature tensor comprises: Performing feature coupling processing on the first target feature tensor and the second target feature tensor respectively to obtain a first coupling feature and a second coupling feature; Performing deformation and transposition processing on the first coupling feature to obtain a fifth sub-feature tensor and a sixth sub-feature tensor; Multiplying the fifth sub-feature tensor and the sixth sub-feature tensor to obtain a third target positioning sub-matrix; Using a normalized exponential function to normalize the third target positioning submatrix to obtain a third normalized target positioning submatrix; Multiplying the third normalized target positioning submatrix and the first coupling feature to obtain a third feature mapping matrix, and performing feature tensor reconstruction on the third feature mapping matrix to obtain a third reconstruction matrix; Performing adaptive scaling processing on the third reconstruction matrix to obtain a third adaptive feature tensor, and performing pixel-by-pixel summing processing on the third adaptive feature tensor and the third feature tensor to obtain the third target feature tensor; The second coupling feature is input into the second target positioning submodule to obtain a fourth target feature tensor.

6. The method according to claim 5, characterized in that The step of inputting the second coupling feature into the second target positioning submodule to obtain a fourth target feature tensor includes: Performing deformation and transposition processing on the second coupling feature to obtain a seventh sub-feature tensor and an eighth sub-feature tensor; Multiplying the seventh sub-feature tensor and the eighth sub-feature tensor to obtain a fourth target positioning sub-matrix; Using a normalized exponential function to normalize the fourth target positioning submatrix to obtain a fourth normalized target positioning submatrix; multiplying the fourth normalized target positioning submatrix and the second coupling feature to obtain a fourth feature mapping matrix, and performing feature tensor reconstruction on the fourth feature mapping matrix to obtain a fourth reconstruction matrix; The fourth reconstruction matrix is ​​adaptively processed by a scaling factor to obtain a fourth adaptive feature tensor, and the fourth adaptive feature tensor and the fourth feature tensor are summed pixel by pixel to obtain the fourth target feature tensor.

7. The method according to claim 1, characterized in that Based on the first camouflaged target position, the first feature tensor, the second feature tensor, the third feature tensor, and the fourth feature tensor are respectively input into a target locking module for iterative processing to obtain a second camouflaged target position, including: Performing upsampling processing on the first camouflaged target position to obtain a first sampling result; Performing convolution processing on the first sampling result, the second feature tensor, and the first convolution layer to obtain a first convolution result; Performing convolution processing on the second feature tensor and the second convolution layer to obtain a second convolution result; Performing convolution processing on the second feature tensor and the third convolution layer, and multiplying the convolution processing result by the first adaptive learning coefficient to obtain a third convolution result; Performing convolution processing on the second feature tensor and the fourth convolution layer, and multiplying the convolution processing result by the second adaptive learning coefficient to obtain a fourth convolution result; Performing upsampling processing on the first feature tensor, and multiplying the sampling processing result by the third adaptive learning coefficient to obtain a second sampling result; Adding the first convolution result, the second convolution result, the third convolution result, the fourth convolution result and the second sampling result to obtain a second sub-camouflaged target position; The second feature tensor, the third feature tensor and the fourth feature tensor are sequentially input into the target locking module for iterative processing to obtain a second camouflaged target position.

8. A camouflaged target detection device based on multi-channel coupling and context information, characterized in that: include: A determination module, used to obtain a target image, and determine a feature tensor of the target image according to a target network model, wherein the feature tensor includes: a first feature tensor, a second feature tensor, a third feature tensor, and a fourth feature tensor; An input module, used for inputting the first feature tensor and the second feature tensor into a target positioning module to obtain a first camouflaged target position; The iteration module is used to input the first feature tensor, the second feature tensor, the third feature tensor and the fourth feature tensor into the target locking module for iterative processing based on the first camouflaged target position to obtain a second camouflaged target position.

9. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method according to any one of claims 1 to 7 when executing the executable instructions or computer programs stored in the memory.

10. A computer-readable storage medium storing executable instructions or a computer program, characterized in that: When the executable instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images