A method for detecting small infrared targets

Abstract

The present invention relates to an infrared small target detection method, which mainly solves the technical problems that the existing infrared small target detection methods have low detection performance in complex scenes such as uneven lighting conditions and cluttered backgrounds, or are prone to losing small target information in the deep network. The method of the present invention uses a three-way position attention dense nested network to extract the initial infrared small target features of the image to be detected, uses an attention weighted sum stacked fusion network to unify the size and fuse the features of the initial infrared small target features, and uses a detection module to detect the infrared small target. The method of the present invention overcomes the problems of unclear features of weak and small targets and lack of surface texture information, retains the position information of small targets, and reduces the probability of losing deep target information.

Description

Technical Field

[0001] The invention relates to a target detection method, in particular to an infrared small target detection method. Background Art

[0002] Infrared small target detection has been widely used in the civilian field, for example, in maritime rescue, forest fire prevention and other application scenarios. Compared with visible light detection, infrared detection uses the principle that any object with a temperature above absolute zero can radiate infrared electromagnetic waves outward in complex scenes with uneven light distribution and cluttered backgrounds. It has the advantage of working all day and all weather; compared with radar detection, infrared detection has stronger concealment. However, under long-distance imaging conditions, due to environmental interference such as atmospheric scattering, the target signal strength is weak, the target occupies a small proportion of pixels, the target surface features are not obvious, and the lack of structural and texture information makes infrared small target detection a great challenge.

[0003] In traditional infrared small target detection methods, due to the use of fixed hyperparameters and manual feature adjustment, it is impossible to adapt to complex scenes with different target shapes and sizes. Infrared small target detection methods based on CNN (Convolutional Neural Networks, CNN) are prone to losing small target information in the deep network due to the use of downsampling and pooling operations. At the same time, since different layers focus on different semantic information, it is crucial to make full use of the effective features extracted by each layer to improve the accuracy of target detection; in addition, since small target pixels account for a small proportion, it is extremely important to strengthen the focus on target pixel information.

[0004] Therefore, a method is needed to solve the above problems in order to improve the model's detection capability for weak infrared targets. Summary of the invention

[0005] The purpose of the present invention is to solve the technical problems that the existing infrared small target detection methods have low detection performance in complex scenes such as uneven lighting conditions and cluttered backgrounds, or are prone to losing small target information in the deep layer of the network, and to propose an infrared small target detection method.

[0006] A method for detecting small infrared targets, which is special in that it includes the following steps:

[0007] 1) Use a three-way position attention dense nested network to extract infrared small target features from the image to be detected;

[0008] The three-way position attention dense nested network includes N layers of detection networks, N≥3; the first layer, the second layer to the Nth layer of the detection network include 1, 2...N residual attention feature extraction trunks in turn, and a position attention feature enhancement block is connected between two adjacent residual attention feature extraction trunks in the second layer and above; the residual attention feature extraction trunk is used to extract the residual attention feature and output the feature C'(i,j), C'(i,j) is used as the input target of the position attention feature enhancement block, the position attention feature enhancement block is used to learn the input target position attention weight information, enhance the three-way position attention dense nested network's perception of position information, and output the enhanced C'(i,j); C'(i,j) is element-by-element multiplied with the enhanced C'(i,j) to obtain C"(i,j); wherein i represents the number of layers where the residual attention feature extraction trunk is located, and j represents the number of residual attention feature extraction trunks in the i-th layer of the detection network; i=1, 2...N, j=1, 2...N;

[0009] The input of the first residual attention feature extraction backbone of the Nth layer is the image to be detected;

[0010] The input of the jth residual attention feature extraction backbone of the i-th layer other than the first one of the N-th layer is the second input, or the result of splicing the first input and the third input, or the result of splicing the first input, the second input and the third input; the first input is the previous C" (i, j-1), the second input is the result of dimensionality reduction of the output C' (i+1, j) of the jth residual attention feature extraction backbone of the i+1-th layer, and the third input is the result of upsampling the output C' (i-1, j-1) of the j-1th residual attention feature extraction backbone of the i-1-th layer;

[0011] The dimension reduction includes maximum pooling and down sampling, wherein when j=1, maximum pooling is used to achieve dimension reduction, and when j≠1, down sampling is used to achieve dimension reduction; the output C'(i,i) of the last residual attention feature extraction trunk of each layer is used as the initial infrared small target feature output by the detection network of this layer;

[0012] 2) Input the initial infrared small target features output by each layer of the detection network into the attention weighted sum stacking fusion network, and output the fusion feature out; the attention weighted sum stacking fusion network includes N channel attention modules and weighted sum stacking blocks connected in sequence; each of the channel attention modules includes a cascaded uniform size module and a channel attention feature enhancement block; the N uniform size modules are respectively connected to the output end of the last residual attention feature extraction trunk corresponding to the N-layer detection network;

[0013] 3) After convolution, the fused feature out is input into the detection module for detection, and the detection module outputs the infrared small target feature.

[0014] Furthermore, in step 2, the initial infrared small target features output by each layer of the detection network are input into the attention weighted sum stacking fusion network, and the output fusion feature out is specifically:

[0015] 2.1) N uniform size modules unify the size of the initial infrared small target features output by the N-layer detection network;

[0016] 2.2) N channel attention feature enhancement blocks perform channel attention feature enhancement on the infrared small target features after unified size, and multiply the initial infrared small target features after unified size by the initial infrared small target features after channel attention feature enhancement element by element, and output out i ;

[0017] 2.3) N weighted summation stack blocks are used to calculate the N out i Weighted along the channel dimension, the weight parameter of each layer along the channel dimension is λ i , out i and the corresponding weight parameter λ i Multiply and output the weighted out i _w; stack the weighted results (out1_w, out2_w, ...out i _w…,out N _w); stack and sum, and output the fused feature out.

[0018] Furthermore, in step 1:

[0019] When i ≥ 3, each layer of the detection network also includes i-2 cascade modules, and the cascade modules are position attention feature enhancement blocks;

[0020] Among them, the input of the first cascade module of the i-th layer is the output C'(i,1) of the first residual attention feature extraction trunk of the i-th layer. After position attention feature enhancement, it is output to the third to the last residual attention feature extraction trunk of the i-th layer respectively, and is spliced ​​with other inputs in the corresponding residual attention feature extraction trunk along the channel dimension;

[0021] The input of the second cascade module of the i-th layer is C'(i,2). After position attention feature enhancement, it is output to the fourth to the last residual attention feature extraction trunk of the i-th layer, and is concatenated with other inputs in the corresponding residual attention feature extraction trunk along the channel dimension;

[0022] Similarly, the input of the mth cascade module in the i-th layer is C'(i,m), which is enhanced with position attention features and output to the m+2th to the last residual attention feature extraction backbone in the i-th layer, and concatenated with other inputs in the corresponding residual attention feature extraction backbone along the channel dimension; m = 1, 2...i-2.

[0023] Furthermore, step 2.1 is specifically as follows:

[0024] Among the N uniform-size modules, the uniform-size module connected to the N layers of N residual attention feature extraction backbones maintains the size of the initial infrared small target features of the layer; the remaining N-1 uniform-size modules upsample and perform convolution processing with a 1*1 convolution kernel size on the initial infrared small target features of the corresponding layer, and output infrared small target features of a size consistent with the aforementioned size, so that the output sizes of all uniform-size modules remain consistent.

[0025] Furthermore, in step 1, the specific steps of extracting the residual attention feature and outputting C'(i,j) by the residual attention feature extraction backbone include:

[0026] A1) When i=N and j=1, the image to be detected is used as the input of the residual attention feature extraction trunk; when i≠N and j=1, the second input is used as the input of the residual attention feature extraction trunk; when i≠N and j=2, the first input, the second input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i≠N and j=3~N, the first input, the second input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=N and j=2, the first input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=N and j=3~N, the first input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk;

[0027] A2) The input of the residual attention feature extraction backbone is successively subjected to convolution processing with a 3*3 convolution kernel size and batch normalization operations;

[0028] A3) After batch normalization, the relu function is used for nonlinear activation operation;

[0029] A4) After the nonlinear activation operation, convolution processing with a convolution kernel size of 3*3 and batch normalization operations are performed in sequence;

[0030] A5) The output of the batch normalization operation enters the channel attention feature enhancement block to perform channel attention feature enhancement and obtain the attention weight;

[0031] A6) multiply the output of the batch normalization operation in step A4 by the attention weights obtained in step A5 element by element;

[0032] A7) performs a residual connection on the input of the residual attention feature extraction backbone of step A1 and the result of element-by-element multiplication of step A6, and then uses a sigmoid function for nonlinear activation operation, and finally outputs C'(i,j).

[0033] Furthermore, in step 2 and step A5, the enhancement step of the channel attention feature enhancement block specifically includes:

[0034] B1) The input is connected in parallel for average pooling and maximum pooling operations;

[0035] B2) The results of average pooling and maximum pooling are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1, and then the relu function is used for nonlinear activation operation;

[0036] B3) The results after the nonlinear activation operation are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1;

[0037] B4) The tensors after the parallel convolution processing in step B3 are added element by element, and then the sigmoid function is used for nonlinear activation operation to obtain the attention weight of each channel to achieve feature adaptive enhancement.

[0038] Furthermore, in step 1, the position attention feature enhancement block enhancement step specifically includes:

[0039] C1) first performs convolution processing on the input C'(i,j) with a convolution kernel size of 1*1, and then uses the relu function for nonlinear activation operation;

[0040] C2) performs convolution processing with a 1*1 convolution kernel size again, and then uses the sigmoid function for nonlinear activation operation to obtain the weight of each pixel and output the enhanced C'(i,j).

[0041] Furthermore, in step 3, the fused feature out is subjected to convolution processing with a convolution kernel size of 1*1; and the detection module adopts an eight-connected area clustering method.

[0042] Furthermore, in step 1, N=5; in step 2.3, λ i =0.2.

[0043] Furthermore, the batch normalization operation adopts a Batch Normalization function.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The method of the present invention uses a three-way position attention dense nested network to extract the initial infrared small target features of the image to be detected, uses an attention weighted sum stacking fusion network to unify the size and fuse the features of the initial infrared small target, and uses a detection module to detect the infrared small target. The method of the present invention overcomes the problems of unclear features of weak small targets and lack of surface texture information, retains the position information of small targets, and reduces the probability of loss of deep target information.

[0046] 2. The method of the present invention uses the position attention feature enhancement block to focus on the important pixel information of the small target on the output of the residual attention feature extraction trunk, and suppresses unimportant pixel points. Through the channel attention feature enhancement block, the output of the uniform size module and the features of the residual attention feature extraction trunk in the extraction process focus on the important channel information of the small target, suppressing unimportant channel information, reducing the semantic gap between different layers, and realizing feature adaptive enhancement.

[0047] 3. The method of the present invention enhances the ability to focus on target pixel position information, enhances the nonlinear characterization ability of the model, improves the model's ability to detect weak infrared targets, and reduces the model's sensitivity to noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of an embodiment of an infrared small target detection method of the present invention (omitting the element-by-element multiplication structure in the three-way position attention dense nested network);

[0049] Figure 2 It is a structural block diagram of a three-way position attention dense nested network in an embodiment of an infrared small target detection method of the present invention (omitting the element-by-element multiplication structure);

[0050] Figure 3 It is a structural block diagram of an attention weighted summation stacking fusion network in an embodiment of an infrared small target detection method of the present invention;

[0051] Figure 4 It is a structural block diagram of an element-by-element multiplication structure in a three-way position attention dense nested network in an embodiment of an infrared small target detection method of the present invention;

[0052] Figure 5 It is a structural block diagram of the residual attention feature extraction backbone in an infrared small target detection method embodiment of the present invention;

[0053] Figure 6 It is a structural block diagram of a channel attention feature enhancement block in an embodiment of an infrared small target detection method of the present invention;

[0054] Figure 7It is a structural block diagram of a position attention feature enhancement block in an embodiment of an infrared small target detection method of the present invention;

[0055] Figure 8 This is a visualization comparison diagram of the detection method of the present invention and other infrared image detection methods;

[0056] Fig. 9 for Figure 8 Corresponding 3D visualization comparison chart;

[0057] Fig.10 Another set of visualization comparison diagrams of the detection method of the present invention and other infrared image detection methods;

[0058] Fig.11 for Fig.10 Corresponding 3D visualization comparison chart. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] The present invention provides an infrared small target detection method, comprising the following steps:

[0061] 1) Use a three-way position attention dense nested network to extract infrared small target features from the image to be detected;

[0062] The three-way position attention dense nested network includes N layers of detection networks, N ≥ 3; Figure 1 , Figure 2 As shown, in this embodiment, N=5. The 1st, 2nd to 5th detection networks include 1, 2...5 residual attention feature extraction backbones respectively. A position attention feature enhancement block (Pixel Attention) is connected between two adjacent residual attention feature extraction backbones (Resnet-Channel AttentionModel, RCA) in the 2nd to 4th detection networks; the jth residual attention feature extraction backbone of the i-th layer is used to extract residual attention features and output features C'(i,j), C'(i,j) is used as the input target of the position attention feature enhancement block, the position attention feature enhancement block connected to the output end of the jth residual attention feature extraction backbone of the i-th layer is used to learn the input target position attention weight information, the three-way position attention densely nested network has the ability to perceive position information, and outputs the enhanced C'(i,j). As shown Figure 4 As shown, C'(i,j) is multiplied element by element with the enhanced C'(i,j) to obtain C"(i,j), thereby enhancing the model's perception of position information. Among them, i represents the layer number of the residual attention feature extraction trunk, and j represents the number of residual attention feature extraction trunks in the i-th layer detection network; i = 1, 2...5, j = 1, 2...5;

[0063] The input of the first residual attention feature extraction backbone in the fifth layer is the image to be detected;

[0064] The input of the jth residual attention feature extraction trunk of the i-th layer other than the first one of the 5th layer is the second input, or the result of splicing the first input and the third input along the channel dimension, or the result of splicing the first input, the second input and the third input along the channel dimension; the first input is the previous C" (i, j-1), the second input is the result of dimension reduction of the output C' (i+1, j) of the jth residual attention feature extraction trunk of the i+1 layer, and the third input is the result of upsampling the output C' (i-1, j-1) of the j-1th residual attention feature extraction trunk of the i-1 layer; dimension reduction includes maximum pooling and downsampling, wherein, when j=1, maximum pooling is used to achieve dimension reduction; when j≠1, downsampling is used to achieve dimension reduction; when j=i, that is, the output C' (i, i) of the last residual attention feature extraction trunk of each layer of the detection network is used as the initial infrared small target feature output by the detection network of this layer;

[0065] Specifically, when i≠5 and j=1, the input of the first residual attention feature extraction backbone of the i-th layer is the second input, that is, C'(i+1,1) after the maximum pooling process. When i=2, 3, 4, and j≠1, the input of the j-th residual attention feature extraction backbone of the i-th layer is the result of the concatenation of the first input C"(i,j-1), the second input downsampled C'(i+1,j) and the third input upsampled C'(i-1,j-1); when i=5 and j≠1, the input of the j-th residual attention feature extraction backbone of the 5th layer is the result of the concatenation of the first input C"(5,j-1) and the third input upsampled C'(4,j-1); the input of the first residual attention feature extraction backbone of the 1st layer is the image to be detected;

[0066] When i ≥ 3, each layer of the detection network also includes i-2 cascade modules, and the cascade modules are position attention feature enhancement blocks; among them, the input of the first cascade module of the i-th layer is the output C'(i,1) of the first residual attention feature extraction trunk of the i-th layer. After the position attention feature is enhanced, it is output to the third to the last residual attention feature extraction trunk of the i-th layer respectively, and is spliced ​​with other inputs of the corresponding residual attention feature extraction trunk along the channel dimension;

[0067] The input of the second cascade module of the i-th layer is C'(i,2). After position attention feature enhancement, it is output to the fourth to the last residual attention feature extraction trunk of the i-th layer, and is spliced ​​with other inputs of the corresponding residual attention feature extraction trunk along the channel dimension;

[0068] Similarly, the input of the mth cascade module in the i-th layer is C'(i,m), which is enhanced with position attention features and output to the m+2th to the last residual attention feature extraction backbone in the i-th layer, and concatenated with other inputs in the corresponding residual attention feature extraction backbone along the channel dimension; m = 1, 2...i-2.

[0069] like Figure 5 As shown in Figure 2, the specific steps of residual attention feature extraction are as follows:

[0070] A1) When i=5 and j=1, the image to be detected is used as the input of the residual attention feature extraction trunk; when i≠5 and j=1, the second input is used as the input of the residual attention feature extraction trunk; when i≠5 and j=2, the first input, the second input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i≠5 and j=3~5, the first input, the second input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=5 and j=2, the first input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=5 and j=3~5, the first input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk;

[0071] A2) The input of the residual attention feature extraction backbone is successively subjected to convolution processing with a 3*3 convolution kernel size and batch normalization (bn) operation;

[0072] A3) After batch normalization, the relu function is used for nonlinear activation operation;

[0073] A4) After the nonlinear activation operation, convolution processing with a convolution kernel size of 3*3 and batch normalization operations are performed in sequence;

[0074] A5) The output of the batch normalization operation enters the channel attention feature enhancement block to perform channel attention feature enhancement;

[0075] like Figure 6 As shown, the specific enhancement steps of the channel attention feature enhancement block include:

[0076] B1) The output of the batch normalization operation is parallelized to perform an average pooling operation and a maximum pooling operation;

[0077] B2) The results of average pooling and maximum pooling are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1, and then the relu function is used for nonlinear activation operation;

[0078] B3) The results after the nonlinear activation operation are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1;

[0079] B4) The results of the parallel convolution processing in step B3 are added element by element, and the tensor after element-by-element addition is nonlinearly activated using the sigmoid function to output the attention weight of the channel attention feature enhancement block. The attention weight is used to weight the corresponding channels, focus on important channels, suppress unimportant channel information, and achieve feature adaptive enhancement.

[0080] A6) multiply the output of the batch normalization operation in step A4 by the attention weights in step A5 element-wise;

[0081] A7) performs a residual connection on the input of the residual attention feature extraction backbone of step A1 and the result of element-by-element multiplication of step A6. After the residual connection, a nonlinear activation operation is performed using a sigmoid function, and finally outputs C'(i,j).

[0082] like Figure 7 As shown, the specific enhancement steps of the position attention feature enhancement block include:

[0083] C1) first performs convolution processing on the input C'(i,j) with a convolution kernel size of 1*1, and then uses the relu function for nonlinear activation operation;

[0084] C2) performs convolution processing with a 1*1 convolution kernel size again, and then uses the sigmoid function for nonlinear activation operation to obtain the weight of each pixel, and finally outputs the enhanced C'(i,j).

[0085] The output of the last residual attention feature extraction trunk of each layer of the detection network is the initial infrared small target feature of that layer;

[0086] 2) Input the initial infrared small target features output by each layer of the detection network into the attention weighted sum stacking fusion network, and output the fusion feature out; a total of N attention weighted sum stacking fusion networks are set, each attention weighted sum stacking fusion network includes a channel attention module and a weighted sum stacking block connected in sequence; the channel attention module includes a cascaded uniform size module and a channel attention feature enhancement block; the N uniform size modules are respectively connected to the output end of the last residual attention feature extraction backbone corresponding to the N layer detection network; Figure 3 As shown, step 2 is specifically as follows:

[0087] 2.1) N uniform size modules unify the size of the initial infrared small target features output by the N-layer detection network;

[0088] In this embodiment, the image to be detected is a feature map of 3*H*W, and the initial infrared small target feature output by the 5th layer detection network is a feature map of 16*H*W; the uniform size module connected to the 5 layers of 5 residual attention feature extraction trunks maintains the size of the initial infrared small target feature of this layer;

[0089] The initial infrared small target features output by the 4th layer detection network are feature maps of 32*H / 2*W / 2; the initial infrared small target features output by the 3rd layer detection network are feature maps of 64*H / 4*W / 4; the initial infrared small target features output by the 2nd layer detection network are feature maps of 128*H / 8*W / 8; the initial infrared small target features output by the 1st layer detection network are feature maps of 256*H / 16*W / 16; the uniform size modules in the 1st to 4th layers upsample the initial infrared small target features of the corresponding layer and perform convolution processing with a convolution kernel size of 1*1, so that the output feature maps have a uniform size of 16*H*W.

[0090] 2.2) N channel attention feature enhancement blocks perform channel attention feature enhancement on the infrared small target features after unified size, and multiply the initial infrared small target features after unified size by the initial infrared small target features after channel attention feature enhancement element by element, and output out i , where i = 1, 2…N;

[0091] 2.3) Each weighted summation stacking block includes a weighting module, a stacking module and a summation module, and the specific steps are as follows:

[0092] 2.3.1) N weighted modules for N out i Weighted along the channel dimension, the weight parameter of each weighted module along the channel dimension is λ i , In this embodiment, λ i =0.2;out i and the corresponding weight parameter λ i Multiply and output the weighted out i _w;

[0093] 2.3.2) The stacking module performs a stacking operation on the weighted results (out1_w, out2_w, ...out i _w…,out N _w);

[0094] 2.3.3) The summation module sums the stacked data and outputs the fused feature out;

[0095] 3) The fused feature out is passed through a convolution layer with a convolution kernel size of 1*1 and input into the detection module for detection. The detection module outputs the infrared small target feature.

[0096] like Figure 8-Figure 11 As shown in the figure, two different groups of images to be detected are obtained after the method of the present invention (ours column) and other infrared image detection methods, and the detection visualization comparison diagram and the corresponding 3D visualization comparison diagram are obtained. Among them, the detection result of the method of the present invention retains the position information of small targets and has better detection ability.

Claims

1. A method for detecting small infrared targets, characterized in that: The following steps are involved: 1) Use a three-way position attention dense nested network to extract infrared small target features from the image to be detected; The three-way position attention dense nested network includes N layers of detection networks, N≥3; the first layer, the second layer to the Nth layer of the detection network include 1, 2...N residual attention feature extraction trunks in turn, and a position attention feature enhancement block is connected between two adjacent residual attention feature extraction trunks in the second layer and above; the residual attention feature extraction trunk is used to extract the residual attention feature and output the feature initial C'(i,j), the initial C'(i,j) is used as the input target of the position attention feature enhancement block, the position attention feature enhancement block is used to learn the input target position attention weight information, enhance the three-way position attention dense nested network's perception of position information, and output the enhanced C'(i,j); the initial C'(i,j) is element-by-element multiplied with the enhanced C'(i,j) to obtain C"(i,j); wherein i represents the number of layers where the residual attention feature extraction trunk is located, and j represents the number of residual attention feature extraction trunks in the i-th layer of the detection network; i=1, 2...N, j=1, 2...N; The input of the first residual attention feature extraction backbone of the Nth layer is the image to be detected; The input of the jth residual attention feature extraction backbone of the i-th layer other than the first one of the N-th layer is the second input, or the result of splicing the first input and the third input, or the result of splicing the first input, the second input and the third input; the first input is the previous C" (i, j-1), the second input is the result of dimensionality reduction of the output initial C' (i+1, j) of the jth residual attention feature extraction backbone of the i+1-th layer, and the third input is the result of upsampling the output initial C' (i-1, j-1) of the j-1th residual attention feature extraction backbone of the i-1-th layer; The dimension reduction includes maximum pooling and down sampling, wherein when j=1, maximum pooling is used to achieve dimension reduction, and when j≠1, down sampling is used to achieve dimension reduction; the output initial C'(i,i) of the last residual attention feature extraction trunk of each layer is used as the initial infrared small target feature output by the detection network of this layer; 2) Input the initial infrared small target features output by each layer of the detection network into the attention weighted sum stacking fusion network, and output the fusion feature out; the attention weighted sum stacking fusion network includes N channel attention modules and weighted sum stacking blocks connected in sequence; each of the channel attention modules includes a cascaded uniform size module and a channel attention feature enhancement block; the N uniform size modules are respectively connected to the output end of the last residual attention feature extraction trunk corresponding to the N-layer detection network; 3) After convolution, the fused feature out is input into the detection module for detection, and the detection module outputs the infrared small target feature.

2. The infrared small target detection method according to claim 1, characterized in that: In step 2, the initial infrared small target features output by each layer of the detection network are input into the attention weighted sum stack fusion network, and the output fusion feature out is specifically: 2.1) N uniform size modules unify the size of the initial infrared small target features output by the N-layer detection network; 2.2) N channel attention feature enhancement blocks perform channel attention feature enhancement on the infrared small target features after unified size, and multiply the initial infrared small target features after unified size by the initial infrared small target features after channel attention feature enhancement element by element, and output out i ; 2.3) N weighted summation stack blocks are used to calculate the N out i Weighted along the channel dimension, the weight parameter of each layer along the channel dimension is λ i , out i and the corresponding weight parameter λ i Multiply and output the weighted out i _w; stack the weighted results (out1_w, out2_w, ...out i _w…,out N _w); stack and sum, and output the fused feature out.

3. The infrared small target detection method according to claim 1, characterized in that: In step 1: When i ≥ 3, each layer of the detection network also includes i-2 cascade modules, and the cascade modules are position attention feature enhancement blocks; Among them, the input of the first cascade module of the i-th layer is the initial output C'(i,1) of the first residual attention feature extraction trunk of the i-th layer. After position attention feature enhancement, it is output to the third to the last residual attention feature extraction trunk of the i-th layer respectively, and spliced ​​with other inputs in the corresponding residual attention feature extraction trunk along the channel dimension; The input of the second cascade module of the i-th layer is the initial C'(i,2). After position attention feature enhancement, it is output to the fourth to the last residual attention feature extraction trunk of the i-th layer, and is concatenated with other inputs in the corresponding residual attention feature extraction trunk along the channel dimension; Similarly, the input of the mth cascade module in the i-th layer is the initial C'(i,m), which is enhanced with position attention features and output to the m+2th to the last residual attention feature extraction backbone in the i-th layer respectively, and concatenated with other inputs in the corresponding residual attention feature extraction backbone along the channel dimension; m=1, 2…i-2.

4. The infrared small target detection method according to claim 2, characterized in that: Step 2.1 is as follows: Among the N uniform-size modules, the uniform-size module connected to the N layers of N residual attention feature extraction backbones maintains the size of the initial infrared small target features of the layer; the remaining N-1 uniform-size modules upsample and perform convolution processing with a 1*1 convolution kernel size on the initial infrared small target features of the corresponding layer, and output infrared small target features of a size consistent with the aforementioned size, so that the output sizes of all uniform-size modules remain consistent.

5. The infrared small target detection method according to claim 3, characterized in that: In step 1, the specific steps of the residual attention feature extraction backbone to extract the residual attention features and output the initial C'(i,j) include: A1) When i=N and j=1, the image to be detected is used as the input of the residual attention feature extraction trunk; when i≠N and j=1, the second input is used as the input of the residual attention feature extraction trunk; when i≠N and j=2, the first input, the second input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i≠N and j=3~N, the first input, the second input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=N and j=2, the first input and the third input are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; when i=N and j=3~N, the first input, the third input and the output of the corresponding cascade module are spliced ​​along the channel dimension as the input of the residual attention feature extraction trunk; A2) The input of the residual attention feature extraction backbone is successively subjected to convolution processing with a 3*3 convolution kernel size and batch normalization operations; A3) After batch normalization, the relu function is used for nonlinear activation operation; A4) After the nonlinear activation operation, convolution processing with a convolution kernel size of 3*3 and batch normalization operations are performed in sequence; A5) The output of the batch normalization operation enters the channel attention feature enhancement block to perform channel attention feature enhancement and obtain the attention weight; A6) multiply the output of the batch normalization operation in step A4 by the attention weights obtained in step A5 element-wise; A7) performs a residual connection between the input of the residual attention feature extraction backbone of step A1 and the result of element-by-element multiplication of step A6, and then uses a sigmoid function for nonlinear activation operation, and finally outputs the initial C'(i,j).

6. The infrared small target detection method according to claim 5, characterized in that: In step 2 and step A5, the enhancement steps of the channel attention feature enhancement block specifically include: B1) The input is connected in parallel for average pooling and maximum pooling operations; B2) The results of average pooling and maximum pooling are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1, and then the relu function is used for nonlinear activation operation; B3) The results after the nonlinear activation operation are processed in parallel by a shared convolution layer with a convolution kernel size of 1*1; B4) The tensors after the parallel convolution processing in step B3 are added element by element, and then the sigmoid function is used for nonlinear activation operation to obtain the attention weight of each channel to achieve feature adaptive enhancement.

7. The infrared small target detection method according to claim 6, characterized in that: In step 1, the position attention feature enhancement block enhancement step specifically includes: C1) The initial input C'(i,j) is first convolved with a 1*1 convolution kernel size, and then the relu function is used for nonlinear activation operation; C2) performs convolution processing with a 1*1 convolution kernel size again, and then uses the sigmoid function for nonlinear activation operation to obtain the weight of each pixel and output the enhanced C'(i,j).

8. The infrared small target detection method according to claim 1, characterized in that: In step 3, the fused feature out is convolved with a convolution kernel size of 1*1; and the detection module adopts an eight-connected area clustering method.

9. The infrared small target detection method according to claim 2, characterized in that: In step 1, N=5; in step 2.3, λ i =0.

2.

10. The infrared small target detection method according to claim 5, characterized in that: The batch normalization operation uses the Batch Normalization function.

Images