Infrared small target detection method based on nested dense fusion and pixel-by-pixel attention

Abstract

The application discloses an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention, and relates to the technical field of infrared detection, and comprises the following steps: acquiring an original image; extracting encoded feature maps of the original image at different scales according to an encoder of a nested U-shaped network; performing dense fusion on the encoded feature maps at different scales to form a first feature map, and processing the first feature map by using a pixel-by-pixel attention mechanism module to obtain an enhanced feature map; performing dense fusion on the enhanced feature map and the first feature map to form a second feature map; decoding the second feature map at different scales according to a decoder of the nested U-shaped network to form decoded feature maps, splicing the decoded feature maps at different scales, and performing convolution operation on the spliced decoded feature maps to obtain a segmentation detection result. The application can effectively suppress background interference and improve the infrared small target detection performance.

Description

Technical Field

[0001] This invention belongs to the field of infrared detection technology, specifically relating to an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention. Background Technology

[0002] Infrared small target detection technology is a key technology in infrared search and tracking systems, and the performance of infrared search and tracking systems depends on the detection accuracy of infrared small targets.

[0003] In existing technologies, the difficulties faced by infrared small target detection in practical applications are: (1) small target size, because the imaging distance is long or the object itself is small, the number of pixels of small targets in infrared images is generally less than 9×9; (2) dark target brightness, or low signal-to-clutter ratio in the image; (3) few target details, because of the principle of infrared thermal imaging, small targets often lack texture information and the edge details are relatively blurry; (4) complex background interference, infrared small targets are easily interfered with by complex backgrounds such as strong clouds, mountains, forests, buildings, and sea clutter; the above difficulties seriously restrict the application of infrared small target detection in practical scenarios.

[0004] Therefore, it is urgent to overcome the shortcomings of existing technologies and improve the accuracy of infrared small target detection. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention, comprising:

[0007] Acquire the raw image, which includes small infrared targets;

[0008] Based on the encoder of the nested U-shaped network, coded feature maps are extracted from the original image at different scales;

[0009] Dense fusion of encoded feature maps at different scales is performed to form a first feature map, which is then processed using a pixel-wise attention mechanism module to obtain an enhanced feature map; the pixel-wise attention mechanism module includes both spatial and channel dimensions.

[0010] The enhanced feature map and the first feature map are densely fused to form the second feature map;

[0011] Based on the decoder of the nested U-shaped network, the second feature map is decoded at different scales to form a decoded feature map. The decoded feature maps at different scales are then concatenated, and the concatenated decoded feature map is then convolved to obtain the segmentation detection result.

[0012] The beneficial effects of this invention are:

[0013] (1) The present invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention. Based on the nested U-shaped network structure, it utilizes jump connections to combine shallow fine-grained information and deep semantic information, which can effectively capture and extract fine-grained detail information of targets in complex backgrounds and has good detection accuracy.

[0014] (2) The present invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention. It constructs a spatial and channel dual-dimensional pixel-by-pixel attention mechanism, which enhances small target features in a fine-grained manner, suppresses background interference, and improves small target detection performance through refined feature attention.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention provided in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of a nested U-shaped network provided in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of a pixel-by-pixel attention mechanism module provided in an embodiment of the present invention;

[0019] Figure 4 (a) is an observation image containing small infrared targets provided in an embodiment of the present invention;

[0020] Figure 4 (b) is another observation image containing small infrared targets provided in an embodiment of the present invention;

[0021] Figure 4 (c) is another observation method including small infrared targets provided in the embodiments of the present invention;

[0022] Figure 4 (d) is another observation image containing small infrared targets provided in an embodiment of the present invention;

[0023] Figure 5 (a) is provided by the embodiments of the present invention and Figure 4 (a) A real label map corresponding to the observation map;

[0024] Figure 5 (b) is provided by the embodiments of the present invention and Figure 4 (b) A real label map corresponding to the observation map;

[0025] Figure 5 (c) is provided by the embodiments of the present invention and Figure 4 (c) A real label map corresponding to the observation map;

[0026] Figure 5 (d) is the embodiment of the present invention and Figure 4 (d) A real label map corresponding to the observation map;

[0027] Figure 6 (a) is provided by the embodiments of the present invention and Figure 4 (a) A schematic diagram of the detection results corresponding to the observation map;

[0028] Figure 6 (b) is provided by the embodiments of the present invention and Figure 4 (b) A schematic diagram of the detection results corresponding to the observation map;

[0029] Figure 6 (c) is provided by the embodiments of the present invention and Figure 4 (c) A schematic diagram of the detection results corresponding to the observation map;

[0030] Figure 6 (d) is provided by the embodiments of the present invention and Figure 4 (d) A schematic diagram of the detection results corresponding to the observation map. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0032] In existing technologies, methods for infrared small target detection can be broadly classified into two categories: traditional methods based on manual modeling and data-driven deep learning methods. Traditional model-driven methods are susceptible to complex backgrounds and noise, are sensitive to parameters, and exhibit poor robustness. Furthermore, low-rank modeling methods require multiple iterations for optimization, and their real-time performance is difficult to match that of some deep learning methods. Data-driven deep learning methods, due to their adaptive feature learning capabilities, are widely used in infrared small target detection. Convolutional neural network-based methods can learn the essential features of targets from large amounts of data, overcoming the limitations of traditional manual modeling. However, existing infrared small target detection methods are often improvements on general target detection methods, such as the method proposed by McIntosh et al. (McIntosh B, Venkataramanan S, Mahalanobis A. Infrared target detection in cluttered environments by maximization of a target to clutter ratio (TCR) metric using a convolutional neural network[J]. IEEE Transactions on Aerospace and ElectronicSystems, 2020, 57(1): 485-496.) is an adjustment within the general target detection framework. It designs a target clutter ratio index based on Faster R-CNN and YOLO-v3, and optimizes the existing general target detection method with this as the target. However, it does not build a targeted network architecture based on the characteristics of infrared small targets.

[0033] In view of this, the present invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention, which detects small targets from infrared images containing complex backgrounds and noise, and solves the problem of stable detection of infrared small targets in strong noise and complex backgrounds.

[0034] Please see Figures 1-2 As shown, Figure 1 This is a flowchart of an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a nested U-shaped network provided in an embodiment of the present invention. The present invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention, comprising:

[0035] S101. Acquire the original image, which includes small infrared targets.

[0036] Specifically, in this embodiment, the original image is first acquired. The original image is an infrared image and includes small infrared targets. Optionally, the original image can be acquired by an outer infrared thermal imaging instrument.

[0037] It should be noted that the original image includes not only small infrared targets but also background interference information.

[0038] S102. Based on the encoder of the nested U-shaped network, extract coded feature maps from the original image at different scales.

[0039] For details, please continue to see Figure 2 As shown, in this embodiment, the nested U-shaped network is a multi-layer network structure, including one network module layer and one transition network module layer. Each network module layer includes one encoder, one decoder, and at most I-1 pixel-wise attention mechanism modules. Optionally, the number of pixel-wise attention mechanism modules in adjacent network module layers differs by 1, and the first network module layer includes I-1 pixel-wise attention mechanism modules. In this embodiment, the nested U-shaped network structure, utilizing jump connections to combine shallow fine-grained information and deep semantic information, can effectively capture and extract fine-grained detail information of targets in complex backgrounds, achieving good detection accuracy.

[0040] In one specific embodiment of the present invention, the nested U-shaped network includes four network module layers and one transition network module layer, consisting of... Figure 2 As shown, the first column on the left is the encoder, the bottom layer is the transition network module layer, and the rightmost column is the decoder. The encoder, transition network module layer, and decoder are all residual U-shaped blocks. The encoder in the first layer is used to extract features from the original image, and the encoders in the second, third, and fourth layers are used to extract features at different scales. The extracted encoded feature maps are passed from the shallow layers to the deep layers to obtain multi-scale encoded feature maps.

[0041] S103. Densely fuse the encoded feature maps of different scales to form a first feature map, and process the first feature map using a pixel-wise attention mechanism module to obtain an enhanced feature map; wherein, the pixel-wise attention mechanism module includes spatial and channel dimensions.

[0042] Specifically, in this embodiment, in order to fully preserve the feature information of infrared small targets, the encoded feature maps are densely fused to form a first feature map.

[0043] It should be noted that dense fusion features include two types:

[0044] Firstly, when the encoder of a nested U-shaped network extracts features, the deep encoder extracts features at different scales (i.e., j=0). The dense fusion node needs to receive the output of the upper-level dense fusion node. For example, the input node of the encoder in the second-layer network module needs to densely fuse the output of the encoder in the first-layer network module. The expression for this dense fusion is:

[0045] L(i,j) = F(Down(L(i-1,j)));

[0046] Where L(i,j) is the feature map output by the j-th dense fusion node in the i-th layer, F() is the multi-layer convolution operation, Down() is the downsampling operation, and the downsampling rate can be 2; Optionally, in this embodiment, the network module layer is set to 4 layers, i = 1, 2, 3, 4, j = 0, 1, 2, 3, 4.

[0047] Secondly, when extracting features in the pixel-by-pixel attention mechanism module, the dense fusion node input to the pixel-by-pixel attention mechanism module needs to receive the outputs of the upper-layer dense fusion node, the lower-layer dense fusion node, and the same-layer jump dense fusion node, i.e., j > 0. The expression for its dense fusion is:

[0048]

[0049] Here, Up() is the upsampling operation, the upsampling rate can be 2, [,] is the concatenation operation, and k is the number of columns in the concatenation process.

[0050] It should be noted that the dense fusion node between the encoder and the pixel-by-pixel attention mechanism module does not need to receive the output of the jump dense fusion node; the dense fusion node between adjacent pixel-by-pixel attention mechanism modules needs to receive the output of both the dense fusion node at the same layer and the output of the jump dense fusion node. By densely fusing multi-scale features, semantic differences can be mitigated, feature information of small infrared targets can be fully preserved, and the false negative rate of targets can be reduced.

[0051] It should also be noted that, please continue to see... Figure 2 As shown, this embodiment also includes a convolution module. All feature maps need to be processed by the convolution module first, and then by the pixel-by-pixel attention mechanism module. Optionally, at least one convolutional layer is provided in the convolution module. In practical applications, the number of convolutional layers can be set selectively, and this embodiment does not limit it.

[0052] Please see Figure 3 As shown, Figure 3This is a schematic diagram of a pixel-wise attention mechanism module provided in an embodiment of the present invention. The pixel-wise attention mechanism module includes a spatially preserved channel attention module and a channel-preserving spatial attention module.

[0053] The spatially preserved channel attention module includes a depthwise separable convolutional layer and a max pooling layer arranged side by side, a first pointwise convolutional layer, a deconvolutional layer, and a first activation function;

[0054] The channel-preserving spatial attention module consists of a second pointwise convolutional layer, a deformable convolutional layer, a third pointwise convolutional layer, and a second activation function arranged sequentially.

[0055] In this embodiment, considering that the classic attention mechanism is prone to losing small target information due to the excessive compression of its spatial and channel dimension information, a pixel-by-pixel attention mechanism with both spatial and channel dimensions is constructed. By refining feature attention, the small target features are enhanced in a fine-grained manner, background interference is suppressed, and the small target detection performance is improved.

[0056] The process of processing the first feature map using the pixel-wise attention mechanism module includes:

[0057] S1031. Based on the first feature map, use the depthwise separable convolutional layer in the spatially preserved channel attention module to obtain the feature map. Figure 1 The features are obtained by using the max-pooling layer in the space-preserving channel attention module. Figure 2 Among them, features Figure 1 and characteristics Figure 2 The spatial dimension of each feature map is 1 / n of the spatial dimension of the first feature map; optionally, n = 4.

[0058] S1032, Features Figure 1 and characteristics Figure 2 Element-wise addition is performed, the number of channels is restored using the first pointwise convolutional layer, the spatial size is restored using the deconvolutional layer, and the first activation function is used for processing to enhance the information of small targets in the channel dimension, thus obtaining the third feature map.

[0059] S1033. Multiply the third feature map element-wise with the first feature map to obtain a channel-by-channel pixel-wise enhanced feature map; wherein, the channel-by-channel pixel-wise enhanced feature map Y c The expression for (X) is:

[0060] Y c (X)=X⊙σ1(DeConv(PWConv1(X1+X2)));

[0061] X1 = MP(X);

[0062] X2 = DWConv(X);

[0063] Where X is the first feature map, σ1 is the first activation function, DeConv() is the deconvolution operation, PWConv1() is the first pointwise convolutional layer operation, MP() is the max pooling operation, and DWConv() is the depthwise separable convolution operation.

[0064] S1034. Based on the feature map enhanced pixel by channel, the channel context information is modulated using the second pointwise convolutional layer, and the number of channels is reduced to obtain the fourth feature map; wherein, the number of channels of the fourth feature map is 1 / r of the number of channels of the feature map enhanced pixel by channel; optionally, r = 4;

[0065] S1035. Based on the fourth feature map, a deformable convolutional layer is used to adaptively capture the features of small targets to obtain the fifth feature map; optionally, the kernel size of the deformable convolutional layer is m×m, where m=3;

[0066] S1036. Based on the fifth feature map, the number of channels is recovered using the third pointwise convolutional layer, and then processed using the second activation function to obtain the sixth feature map.

[0067] S1037. Multiply the sixth feature map element-wise with the channel-by-channel pixel-by-pixel enhanced feature map to obtain the enhanced feature map; wherein, the expression of the enhanced feature map is:

[0068] Y(Y c (X))=Y c (X)⊙σ2(PWConv2(DeformConv(Y c (X))));

[0069] PWConv2() is the second pointwise convolutional layer operation, and DeformConv() is the deformable convolutional layer operation.

[0070] S104. Densely fuse the enhanced feature map and the first feature map to form the second feature map.

[0071] For details, please continue to see Figure 2 As shown, in this embodiment, the features of the encoder with the input nested U-shaped network are densely fused to form a multi-scale second feature map. Dense fusion can alleviate semantic differences and fully preserve the feature information of infrared small targets.

[0072] S105. Based on the decoder of the nested U-shaped network, the second feature map is decoded at different scales to form a decoded feature map. The decoded feature maps at different scales are concatenated, and the concatenated decoded feature map is then convolved to obtain the segmentation detection result.

[0073] Specifically, in this embodiment, the convolution operation is performed on the spliced ​​decoded feature map, and the convolution kernel used is 1×1.

[0074] This invention provides an infrared small target detection method based on nested dense fusion and pixel-by-pixel attention. It uses a nested U-shaped network structure and utilizes jump connections to combine shallow fine-grained information and deep semantic information, which can effectively capture and extract fine-grained detail information of targets in complex backgrounds and has good detection accuracy.

[0075] In an optional embodiment of the present invention, the third feature map has the same dimension as the first feature map, and the sixth feature map has the same dimension as the first feature map.

[0076] Specifically, in this embodiment, the dimensions of the third feature map and the sixth feature map are the same as those of the first feature map, which can enhance the ability to represent small target details, avoid the loss of small target information due to compression space and channel dimensions, and improve the detection accuracy of small targets.

[0077] In an optional embodiment of the present invention, the performance of the detection method proposed in this invention is verified through experiments. Figure 4 (a) is an observation image containing small infrared targets provided in an embodiment of the present invention. Figure 4 (b) is another observation image containing small infrared targets provided by an embodiment of the present invention. Figure 4 (c) is another observation image containing small infrared targets provided in an embodiment of the present invention. Figure 4 (d) is another observation image containing small infrared targets provided in an embodiment of the present invention. Figure 5 (a) is provided by the embodiments of the present invention and Figure 4 (a) A true label map corresponding to the observation map. Figure 5 (b) is provided by the embodiments of the present invention and Figure 4 (b) A real label map corresponding to the observation map. Figure 5 (c) is provided by the embodiments of the present invention and Figure 4 (c) A true label map corresponding to the observation map. Figure 5 (d) is provided by the embodiments of the present invention and Figure 4 (d) A true label map corresponding to the observation map. Figure 6 (a) is provided by the embodiments of the present invention and Figure 4 (a) A schematic diagram of the detection results corresponding to the observation map. Figure 6 (b) is provided by the embodiments of the present invention and Figure 4 (b) A schematic diagram of the detection results corresponding to the observation map. Figure 6 (c) is provided by the embodiments of the present invention and Figure 4 (c) A schematic diagram of the detection results corresponding to the observation map. Figure 6(d) is provided by the embodiments of the present invention and Figure 4 (d) A schematic diagram of the detection results corresponding to the observation map; from Figure 4 It is evident that infrared small targets, in scenarios with complex background interference and significant clutter, are subject to... Figure 5 and Figure 6 As shown, the detection method provided by this invention can effectively extract small target images that are close to the real labels, thus achieving the purpose of infrared small target detection in complex backgrounds.

[0078] In summary, this invention provides an infrared small target detection method based on nested dense fusion and pixel-wise attention. First, a nested U-shaped network is used to expand the receptive field and extract multi-scale features within a single layer to enrich the feature representation of small-scale targets and enhance the network's ability to distinguish small-scale targets. Second, dense fusion is used to further preserve small target information, alleviate the semantic gap when fusing features from different layers, and reduce the false negative rate. Finally, pixel-wise attention in both spatial and channel dimensions is applied to the fused feature map to enhance small targets and suppress interference from complex backgrounds.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0080] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0081] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting small infrared targets based on nested dense fusion and pixel-by-pixel attention, characterized in that, include: Acquire the raw image, which includes small infrared targets; Based on the encoder of the nested U-shaped network, encoded feature maps are extracted from the original image at different scales; wherein, the nested U-shaped network is a multi-layer network structure, including... The system comprises a layered network module layer and a transitional network module layer, wherein the network module layer includes one encoder, one decoder, and at most... Each pixel-wise attention mechanism module; the number of pixel-wise attention mechanism modules in two adjacent network module layers differs by 1, and the first layer of the network module includes A pixel-wise attention mechanism module; the nested U-shaped network includes 4 network module layers and one transition network module layer. The first column on the left is the encoder, the bottom layer is the transition network module layer, and the rightmost column is the decoder. The encoder, the transition network module layer, and the decoder are all residual U-shaped blocks. The encoded feature maps at different scales are densely fused to form a first feature map, and the first feature map is processed using a pixel-wise attention mechanism module to obtain an enhanced feature map; wherein, the pixel-wise attention mechanism module includes both spatial and channel dimensions; The enhanced feature map and the first feature map are densely fused to form a second feature map; The second feature map is decoded at different scales using a nested U-shaped network decoder to form a decoded feature map. The decoded feature maps at different scales are then concatenated, and the concatenated decoded feature maps are then subjected to a convolution operation to obtain a segmentation detection result.

2. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 1, characterized in that, Also includes: Convolutional module; The first feature map is first processed by the convolution module, and then by the pixel-by-pixel attention mechanism module. All the first feature maps input to the convolution module are first densely fused.

3. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 1, characterized in that, The dense fusion process includes: when The dense fusion node accepts the output of the upper-layer dense fusion node, and its expression is: ； in, For the first Layer Feature maps output by dense fusion nodes For multi-layered convolution operations, This is a downsampling operation; when A dense fusion node accepts the outputs of upper-level dense fusion nodes, lower-level dense fusion nodes, and same-level jump dense fusion nodes. Its expression is: ； in, For upsampling operation, For splicing operations, This represents the number of columns in the splicing process.

4. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 1, characterized in that, The pixel-by-pixel attention mechanism module includes a spatially preserved channel attention module and a channel-preserving spatial attention module; The spatially preserved channel attention module includes a depthwise separable convolutional layer and a max pooling layer arranged in parallel, a first pointwise convolutional layer, a deconvolutional layer, and a first activation function; The spatial attention module for channel retention includes a second pointwise convolutional layer, a deformable convolutional layer, a third pointwise convolutional layer, and a second activation function arranged sequentially.

5. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 4, characterized in that, The process of processing the first feature map using a pixel-wise attention mechanism module to obtain an enhanced feature map includes: Based on the first feature map, feature map one is obtained by using a depthwise separable convolutional layer in the spatially preserved channel attention module; feature map two is obtained by using a max pooling layer in the spatially preserved channel attention module; wherein, the spatial dimensions of feature map one and feature map two are both times the spatial dimensions of the first feature map. ; The first feature map and the second feature map are added element by element, the number of channels is restored using the first pointwise convolutional layer, the spatial size is restored using the deconvolutional layer, and the third feature map is obtained by processing using the first activation function. The third feature map is multiplied element-wise with the first feature map to obtain a channel-by-pixel enhanced feature map; Based on the pixel-by-pixel enhanced feature map, a fourth feature map is obtained using the second pointwise convolutional layer; wherein, the number of channels in the fourth feature map is equal to the number of channels in the pixel-by-pixel enhanced feature map. ; Based on the fourth feature map, the deformable convolutional layer is used to adaptively capture the features of small targets to obtain the fifth feature map. Based on the fifth feature map, the number of channels is recovered using the third pointwise convolutional layer, and then processed using the second activation function to obtain the sixth feature map; The sixth feature map is multiplied element-wise with the channel-by-pixel enhanced feature map to obtain the enhanced feature map.

6. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 5, characterized in that, The third feature map has the same dimension as the first feature map, and the sixth feature map has the same dimension as the first feature map.

7. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 5, characterized in that, The feature map of the channel is enhanced pixel by pixel. The expression is: ； ； ； in, This is the first feature map. As the first activation function, For deconvolution operation, This is the first pointwise convolutional layer operation. For max pooling operation, This is a depthwise separable convolution operation.

8. The infrared small target detection method based on nested dense fusion and pixel-by-pixel attention according to claim 5, characterized in that, The expression for the enhanced feature map is: in, This is the operation for the second pointwise convolutional layer. This is a deformable convolutional layer operation.

Images