Infrared small target detection method and system based on Transform and multi-scale cavity convolution spatial pyramid

Abstract

The invention discloses an infrared small target detection method and system based on Transform and a multi-scale cavity convolution spatial pyramid. The method comprises the following steps: obtaining an infrared small target image set of a detected object; an infrared small target detection network model based on a Transform and a multi-scale cavity convolution space pyramid is constructed; and inputting the infrared small target image set of the detected object into the trained infrared small target detection network model based on the Transform and the multi-scale cavity convolution space pyramid to realize small target detection of the detected object. According to the method, the multi-head attention mechanism and the multi-scale cavity convolution spatial pyramid network are utilized, the feature information in the infrared small target image can be effectively utilized, and the detection accuracy and robustness of the small target are improved. Meanwhile, a lightweight convolution structure is adopted, the complexity of the network is reduced, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the field of infrared small target detection, and specifically relates to an infrared small target detection method and system based on Transformer and multi-scale dilated convolutional spatial pyramid. Background Art

[0002] Infrared small target detection technology is a technology that uses infrared sensors to detect and identify targets. Infrared sensors can sense the heat emitted by targets and have a high detection rate and accuracy for some small targets such as people, animals, and vehicles.

[0003] Infrared small target detection technology is mainly used in military reconnaissance, security monitoring, drone navigation, fire warning and other fields. In military reconnaissance, infrared small target detection technology can help the army to find the enemy's hidden targets in time and improve combat efficiency; in security monitoring, it can help the police to find suspicious targets in time and improve the level of public security; in drone navigation, it can help drones avoid obstacles and complete flight missions.

[0004] The core of infrared small target detection technology is to use infrared sensors to collect the heat information of the target, and then detect and identify the target through image processing, pattern recognition and other technologies. Traditional methods for infrared small target detection include using image enhancement techniques such as histogram equalization and filtering to improve the contrast between the target and the background for subsequent processing. There are also background prediction and subtraction methods to suppress the background, such as high-pass filtering, wavelet transform, and difference method. Some deep learning methods for infrared small target detection are based on neural networks, which can automatically extract image features through training with a large amount of labeled data and are applied to infrared small target detection. Some methods are based on region extraction networks (RPNs), such as Faster R-CNN, which can effectively extract target areas. There are also some single-stage detectors, such as YOLO and SSD, which can achieve real-time detection, but may have low accuracy in small target detection. However, traditional methods rely heavily on the contrast between the target and the background and are difficult to deal with low-contrast scenes. In complex backgrounds, the background suppression effect is not good, which easily leads to missed detection or false detection. Some high-complexity algorithms have poor real-time performance and are not suitable for real-time applications. Neural network-based methods require a large amount of labeled data, and it is difficult to obtain and annotate infrared small target data sets. Deep learning models are not as effective in detecting small targets as large targets, and are prone to miss detection when the target size is small. Deep learning models usually have high computational complexity and high hardware requirements. The generalization ability in different scenarios may be poor, and a large amount of data from different scenarios is required for training. At present, with the development of computer vision, artificial intelligence and other technologies, infrared small target detection technology is constantly improving and has a broader application prospect. Summary of the invention

[0005] In view of the problems of low resolution and contrast of small infrared targets, small targets, and targets often appearing as hot spots or red dots, the present invention proposes a method and system for detecting small infrared targets based on Transformer and multi-scale dilated convolutional spatial pyramid. By using a multi-head attention mechanism and a multi-scale dilated convolutional spatial pyramid network, the feature information in the image of small infrared targets can be effectively utilized to improve the detection accuracy and robustness of small targets. At the same time, a lightweight convolutional structure is adopted to reduce the complexity of the network and improve the computational efficiency. With the continuous development and wide application of infrared imaging technology, the present invention can be widely used in military reconnaissance, security monitoring, drone navigation, fire warning and other fields.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for detecting small infrared targets based on Transformer and multi-scale dilated convolutional spatial pyramid includes the following steps:

[0008] Obtaining a set of infrared small target images of the inspected object;

[0009] Construct an infrared small target detection network model based on Transformer and multi-scale dilated convolutional spatial pyramid;

[0010] The infrared small target image set of the inspected object is input into the trained infrared small target detection network model based on Transformer and multi-scale dilated convolution spatial pyramid to realize small target detection of the inspected object.

[0011] Preferably, the infrared small target image set of the inspected object is:

[0012]

[0013] in, are k detected infrared small target images, and the infrared small target image set I h The total number of elements is K, and the size of the infrared small target image is n ch ×h×w,n ch is the infrared small target image channel, h is the infrared small target image height, and w is the infrared small target image width.

[0014] Preferably, the infrared small target detection network model based on Transformer and multi-scale dilated convolutional spatial pyramid includes: a backbone network with a multi-scale dilated convolutional spatial pyramid module, a neck network and 5 feature detection heads of different sizes;

[0015] Among them, the backbone network includes: a convolutional layer, a convolution and SK attention mixed layer CBSK, a movable window transformer and convolution mixed network layer Swin-Conv and a multi-scale void convolution spatial pyramid layer.

[0016] Preferably, the method of constructing a convolutional and SK attention mixed layer CBSK includes:

[0017] Assume that given an input feature map F∈R C×H×W , first passes through a module consisting of convolution, normalization, and activation functions in series, and the process is expressed as:

[0018] F 1 (F) = SiLu(BN(Conv(F) 1×1 )),

[0019] In the formula, SiLu(·) represents the activation function, BN(·) represents the normalization function, Conv(·) 1×1 represents the convolution function with a convolution kernel of 1, F 1 (·) represents the convolutional normalization activation function;

[0020] Next, the output continues to pass through three modules in series, and the SK attention mechanism is added to the modules in series. The process is expressed as:

[0021] F k (F k-1 )=SK(F 1 (F k-1 ))+F k-1 ,

[0022] Where K = 2, 3, 4, and SK(·) represents the SK attention function.

[0023] Preferably, the movable window transformer and convolution hybrid network layer Swin-Conv includes: starting from a structure layer consisting of convolution, normalization, and activation, the middle part has two series-connected Swin-Transformer modules, and the last component is a convolution and SK attention hybrid layer CBSK module.

[0024] Preferably, the multi-scale atrous convolution spatial pyramid module consists of five branches;

[0025] The first branch uses the residual structure to retain shallow information;

[0026] The remaining branches first extract features through the dilated convolution layer. Among them, the kernel size of the dilated convolution layer in the four branches is 1×1, and the rest are 3×3, with expansion rates of 1, 6, 12, and 18 respectively. Then, the feature enters a structure consisting of a normalization layer and an activation function in series. Finally, the feature is processed by three consecutive maximum pooling layers and a residual structure.

[0027] The outputs of these five branches are concatenated to obtain the final feature.

[0028] Preferably, the infrared small target image set of the inspected object is input into the trained infrared small target detection network model based on Transformer and multi-scale dilated convolution spatial pyramid to realize the method of small target detection of the inspected object, which includes:

[0029] Use size n ch The infrared small target image of size ×h×w is taken as input into the backbone network, and five feature maps of different scales are obtained and output through the convolution layer, the convolution and SK attention mixed layer CBSK, the movable window transformer and convolution mixed network layer Swin-Conv and the multi-scale void convolution spatial pyramid layer. and The scales are and Among them, the number of all Swin-Conv layers is set to 1;

[0030] The five feature maps of different scales and The neck network is input, and after upsampling, concatenation, C2f layer, and convolution layer processing, 5 feature maps of different scales are obtained. 1 , D 2 , D 3 , D 4 and D 5 , the scales are and

[0031] The five feature maps D of different scales are 1 , D 2 , D 3 , D 4 and D 5 The feature detection head is input, and 5 detection heads of different scales are designed in the feature detection head to realize the detection of large targets and small targets.

[0032] The present invention also provides an infrared small target detection system based on Transformer and multi-scale dilated convolutional spatial pyramid, comprising: an acquisition module, a construction module and a detection module;

[0033] The acquisition module is used to obtain a set of infrared small target images of the inspected object;

[0034] The building module is used to build an infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid;

[0035] The detection module is used to input the infrared small target image set of the inspected object into the trained infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid to realize small target detection of the inspected object.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention provides an infrared small target detection method and system based on Transformer and multi-scale dilated convolution spatial pyramid. The Swin-Conv module proposed in the present invention can make maximum use of shallow features, can accurately detect infrared small targets under complex backgrounds, and also has high detection accuracy for unformed and dim infrared small targets. The invention has detection heads of multiple scales and can effectively detect targets of different sizes. The invention can accurately detect infrared small targets under complex backgrounds. The invention has broad application prospects in the field of infrared target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 is a schematic diagram of an infrared imaging system according to an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of the overall architecture of a detection network according to an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of a CBSK module according to an embodiment of the present invention;

[0042] Figure 4 It is a structural diagram of the SK attention mechanism of an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of a Swin-Conv module according to an embodiment of the present invention;

[0044] Figure 6Schematic diagram of a dilated convolutional spatial pyramid module according to an embodiment of the present invention;

[0045] Figure 7 The present invention is a flowchart of an infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid.

[0046] Description of the drawings: 103 - infrared camera 103; 104 - target; 102 - computer; 101 - infrared small target image set after detection. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Embodiment 1

[0050] like Figure 7 As shown, the present invention provides an infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid, including: a computer, an infrared imaging device, and an object to be inspected; the computer is connected to the infrared imaging device. The computer performs small target detection on the infrared small target image of the object to be inspected through the detection method proposed by the present invention. By using an infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid, an infrared small target image is captured by an infrared imaging device, and the infrared target image is detected, and the implementation steps are as follows:

[0051] Obtaining a set of infrared small target images of the inspected object;

[0052] Construct an infrared small target detection network model based on Transformer and multi-scale dilated convolutional spatial pyramid;

[0053] The infrared small target image set of the inspected object is input into the trained infrared small target detection network model based on Transformer and multi-scale void convolution spatial pyramid to realize small target detection of the inspected object.

[0054] In this embodiment, Figure 1It is an infrared imaging system, where the infrared camera 103 photographs the target 104, and after being processed by the computer 102, the detected infrared small target image set 101 is obtained as follows:

[0055]

[0056] in, are k detected infrared small target images, and the infrared small target image set I h The total number of elements is K, and the size of the infrared small target image is n ch ×h×w,n ch is the infrared small target image channel, h is the infrared small target image height, and w is the infrared small target image width.

[0057] In this embodiment, the following Figure 2 The network model shown in the figure, the infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid includes: a backbone network with a multi-scale hole convolution spatial pyramid module, a neck network and 5 feature detection heads of different sizes;

[0058] Among them, the backbone network includes: a convolutional layer, a convolution and SK attention mixed layer CBSK, a movable window transformer and convolution mixed network layer Swin-Conv and a multi-scale void convolution spatial pyramid layer.

[0059] In this embodiment, the structure of CBSK is as follows Figure 3 As shown, the method of constructing the convolution and SK attention mixed layer CBSK includes:

[0060] Assume that given an input feature map F∈R C×HxW (Specifically: F∈R 3×640×640 ), first passes through a module consisting of convolution, normalization, and activation functions in series, and the process can be written as:

[0061] F 1 (F) = SiLu(BN(Conv(F) 1×1 )),

[0062] In the formula, SiLu(·) represents the activation function, BN(·) represents the normalization function, Conv(·) 1×1 represents the convolution function with a convolution kernel of 1, F 1 (·) represents the convolutional normalization activation function;

[0063] Next, the output continues to pass through three modules in series. The invention adds the SK attention mechanism to this module to improve the feature extraction capability. The process can be written as:

[0064] F k (Fk-1 )=SK(F 1 (F k-1 ))+F k-1 ,

[0065] In the formula, K = 2, 3, 4, SK (·) represents the SK attention function. Its structure is shown in the figure Figure 4 As shown. This method is called "selective kernel", which allows each neuron to adaptively adjust the size of its receptive field (convolution kernel) according to the multi-scale of the input information. In this way, the multi-scale features of the complex image space can be better captured while avoiding wasting a lot of computing resources. The SK attention mechanism has another advantage, which is that it can aggregate deep features while improving interpretability. First, given the input feature F 1 ∈R 16×640×640 , first through a convolution with a kernel size of 1×1, the output is obtained X. Then through two convolutions with different kernel sizes, the outputs of the two branches are obtained Y 1 and Y 2 Then, the outputs of the two branches are added together, and then the output Y is obtained through the pooling layer and the fully connected layer. Finally, the output Y is added to Y 1 and Y 2 Perform multiplication and then add to get the final output Z. The process can be written as:

[0066] X=Conv(F 1 ) 1×1

[0067] Y 1 =Conv(X) 3×3

[0068] Y 2 =Conv(X) 5×5

[0069] Y=FC(GP(Y 1 +Y 2 ))

[0070] Z=Y 1 ×Y+Y 2 ×Y

[0071] In the above formula, Conv(·) 3×3 and Conv(·) 5×5 They represent convolution functions with kernel sizes of 3×3 and 5×5, respectively. GP(·) is the global pooling function, FC(·) is the fully connected function, and Z is the final output.

[0072] In this embodiment, the Swin-Conv module is as follows Figure 5As shown in (a), this module can maximize the use of shallow features and can accurately detect small infrared targets in complex backgrounds. It also has high detection accuracy for unshaped and dim infrared small targets. The movable window transformer and convolution hybrid network layer Swin-Conv includes: starting from a structural layer consisting of convolution, normalization, and activation. The middle part of the network has two Swin-Transformer modules connected in series. The last component of the architecture is the CBSK module. Here, the Swin-Transformer block is used to extract deep semantic information while retaining shallow semantic information through the residual structure. The feature images are obtained respectively. and The invention gives an input feature map a∈R 16×H×W (Specifically a∈R 16×640×640 ), each time through a stage, the width and height of the feature map will be reduced by half, and finally 4 feature images of different scales are obtained and (Specifically: 32×320×320, 64×160×160, 128×80×80 and 256×40×40). The process can be expressed as:

[0073] b=SiLu(BN(Conv(a) 1×1 ))

[0074] c=Swin(c)+c

[0075] I k =CBSK(b)

[0076] In the above formula, K = 1, 2, 3, 4, a represents input, I k represents the output of each layer in the backbone network, SiLu(·) represents the activation function, BN(·) represents the normalization function, Conv(·) 1×1 represents a convolution function with a convolution kernel of 1, b and c are intermediate variables, CBSK(·) is the processing function of the CBSK module proposed in the present invention, and Swin(·) is the processing function of the Swin-Transformer module. Its structure is as follows: Figure 5 As shown in (b), its movable window attention mechanism is used to capture local and global context information, and the deep feature map is weighted through the attention mechanism. At the same time, it combines the weighted deep feature map with the input feature map to retain the original feature information. The process can be expressed as:

[0077] H 1 =MLP(LN(WMSA(LN(H))+H)+(WMSA(LN(H))+H)

[0078] Out = MLP (LN (SWMSA (LN (H 1 ))+H 1 )+(SWMSA(LN(H 1 ))+H 1 )

[0079] In the above formula, H represents the input feature, H 1 represents the intermediate variable, MLP(·) is a multi-layer perceptron function, LN(·) represents the layer normalization function, WMSA(·) is a window multi-head attention function, and SWMSA(·) is a moving window multi-head attention function.

[0080] In this embodiment, the structure of the dilated convolutional spatial pyramid module is as follows: Figure 6 As shown. The multi-scale atrous convolution spatial pyramid module (MASPP) consists of five branches. The first branch uses a residual structure to retain shallow information. The remaining branches first extract features through a dilated convolution layer. Among them, the kernel size of the dilated convolution layer in the four branches is 1×1, and the rest are 3×3. Their expansion rates are 1, 6, 12, and 18, respectively. Then, the feature enters a structure consisting of a normalization layer and an activation function in series. Finally, the feature is processed by three consecutive maximum pooling layers and a residual structure. Finally, the outputs of the five branches are spliced ​​to obtain the final feature This module can connect the feature map obtained in the channel dimension with four atrous convolutional layers with different dilation rates. It can solve the alignment problem between anchor points and feature maps. Initially, the input features are sampled at different rates and extracted from different scales to obtain different receptive fields (RFs). The process is expressed as,

[0081] K′=K+(K-1)×(d-1)

[0082] F i+1 =F i +(K′-1)×Str i

[0083]

[0084] In the above formula, K′ is the equivalent convolution kernel of the dilated convolution, K is the kernel size of the normal convolution, and d is the dilation rate of the dilated convolution. i+1 is the receptive field of the current layer. i is the receptive field of the previous layer and Str i Represents the product of the stride length of the previous layer. i Indicates the step size of the i-th layer dilated convolution movement.

[0085] In this embodiment, the infrared small target image set of the detected object is input into the trained infrared small target detection network model based on Transformer and multi-scale dilated convolution spatial pyramid, and the method for realizing small target detection of the detected object includes:

[0086] Use size n ch ×h×w (specifically: 3×640×640) images are taken as input into the backbone network, and 5 feature maps of different scales are obtained and output through the convolution layer, the convolution and SK attention mixed layer CBSK, the movable window transformer and convolution mixed network layer Swin-Conv and the multi-scale void convolution spatial pyramid layer. and The scales are and (Specifically: 32×320×320, 64×160×160, 128×80×80, 256×40×40, and 512×20×20), where the number of all Swin-Conv layers is set to 1;

[0087] The five feature maps of different scales and Enter the neck network, first The intermediate variable X is obtained through the MASPP module 5 . D 5 After upsampling Splicing, and then passing through the C2f layer and the convolution layer to obtain the intermediate variable X 4 . D 4 After upsampling Splicing, and then passing through the C2f layer and the convolution layer to obtain the intermediate variable X 3 . D 3 After upsampling Splicing, and then passing through the C2f layer and the convolution layer to obtain the intermediate variable X 2 . D 2 After upsampling Splicing, and then passing through the C2f layer and the convolution layer to obtain the intermediate variable X 1 .X 1 After passing through the C2f layer and the convolution layer, it is combined with X after passing through the C2f layer 3 Splicing to get D 2 . D 2 After passing through the C2f layer and the convolution layer, it is combined with X after passing through the C2f layer 3 Splicing to get D 3 . D 3 After passing through the C2f layer and the convolution layer, it is combined with X after passing through the C2f layer 4 Splicing to get D4 . D 4 After passing through the C2f layer and the convolution layer, it is combined with X after passing through the C2f layer 5 Splicing to get D 5 .X 1 After passing through the C2f layer and the convolution layer, it is combined with X after passing through the C2f layer 2 Splicing to get D 1 Finally, we get 5 feature maps of different scales D 1 , D 2 , D 3 , D 4 and D 5 , the scales are and (Specifically: 32×320×320, 64×160×160, 128×80×80, 256×40×40 and 512×20×20); it realizes the fusion of feature information and context at different scales and improves the performance of the network.

[0088] The five feature maps D of different scales are 1 , D 2 , D 3 , D 4 and D 5 The feature detection head is input, and five detection heads with scales of 320×320, 160×160, 80×80, 40×40 and 20×20 are designed in the feature detection head to realize the detection of large and small targets.

[0089] In this embodiment, the training is performed in a manner of gradually decreasing the learning rate, and the final learning rate is reduced to 0.001, and the loss function is LOSS. The training parameters are set as follows: the initial learning rate lr is 0.01, the batch size is 16, the training set validation set test set is divided into 7:1:2, the optimizer is SGD, and the total training cycle is 300.

[0090] LOSS=loss cls +loss obj +loss ciou ,

[0091] In the above formula, loss cls is the category loss, loss obj is the confidence loss, loss ciou is the bounding box regression loss calculated using the Complete Intersection over Union (CIoU) loss function.

[0092] In this embodiment, the trained network is predicted to output the target image and the predicted target defect frame. First, the image to be tested Img is input into the network. The size of the image is n×h×w (specifically: 3×640×640). After network processing, the output prediction frames of 5 detection heads are obtained, and the output feature map scales are h / 2×w / 2, h / 4×w / 4, h / 8×w / 8, h / 16×w / 16 and h / 32×w / 32 (specifically: 320×320, 160×160, 80×80, 40×40 and 20×20). Each prediction frame is represented by four coordinate values, and the final prediction frame output is obtained after the post-processing module. The post-processing module mainly consists of two parts, namely the non-maximum suppression module and the box range module. Among them, the non-maximum suppression module first filters out a part through the confidence threshold (IOU=0.6), and after sigmoid, they are all between (0, 1). The maximum value of the two is compared with the threshold, and the XYWH format is converted to XYXY format. Then, the category confidence and label subscript of these are picked out through the category tensor. Finally, an offset is added to the prediction box to complete the filtering of the label box through non-maximum suppression. An offset is added to different categories to distinguish different categories. Finally, a matrix of 3 rows and 6 columns will be obtained, representing the three predicted targets and their corresponding XYXY format boxes, category confidence, and category subscripts. The box range module maps the prediction results back to the original input image size, and crops the obtained XYXY coordinate information to the specified image size range to ensure that the bounding box does not exceed the actual size of the image. In short, the prediction box is not allowed to exceed the original image size.

[0093] If the image to be detected outputs a detection frame, it means that the target is detected in the infrared image to be detected, otherwise, the target is not detected.

[0094] Embodiment 2

[0095] The present invention also provides an infrared small target detection system based on Transformer and multi-scale dilated convolutional spatial pyramid, comprising: an acquisition module, a construction module and a detection module;

[0096] The acquisition module is used to obtain a set of infrared small target images of the inspected object;

[0097] The construction module is used to build an infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid;

[0098] The detection module is used to input the infrared small target image set of the inspected object into the trained infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid to realize small target detection of the inspected object.

[0099] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for infrared small target detection based on Transformer and multi-scale dilated convolutional spatial pyramid, characterized in that: The following steps are involved: Obtaining a set of infrared small target images of the inspected object; Construct an infrared small target detection network model based on Transformer and multi-scale dilated convolutional spatial pyramid; The infrared small target image set of the inspected object is input into the trained infrared small target detection network model based on Transformer and multi-scale dilated convolution spatial pyramid to realize small target detection of the inspected object.

2. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 1 is characterized in that: The infrared small target image set of the inspected object is: in, are k detected infrared small target images, and the infrared small target image set I h The total number of elements is K, and the size of the infrared small target image is n ch ×h×w,n ch is the infrared small target image channel, h is the infrared small target image height, and w is the infrared small target image width.

3. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 2 is characterized in that: The infrared small target detection network model based on Transformer and multi-scale dilated convolutional spatial pyramid includes: a backbone network with a multi-scale dilated convolutional pyramid module, a neck network and 5 feature detection heads of different sizes; The backbone network includes: convolutional layer, convolution and SK attention mixed layer CBSK, movable window transformer and convolutional hybrid network layer Swin-Con v and multi-scale atrous convolutional spatial pyramid layers.

4. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 3 is characterized in that: The method of constructing the convolution and SK attention mixed layer CBSK includes: Assume that given an input feature map F∈R C×H×W , first passes through a module consisting of convolution, normalization, and activation function in series, the process is expressed as: F1(F)=SiLu(BN(Conv(F) 1×1 )), In the formula, SiLu(·) represents the activation function, BN(·) represents the normalization function, Conv(·) 1×1 represents the convolution function with a convolution kernel of 1, and F1(·) represents the convolution normalization activation function; Next, the output continues to pass through three modules in series, and the SK attention mechanism is added to the modules in series. The process is expressed as: F k (F k-1 )=DK(F1(F k-1 ))+F k-1 , Where K = 2, 3, 4, and SK(·) represents the SK attention function.

5. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 3 is characterized in that: Movable Window Transformer and Convolutional Hybrid Network Layer Swin-Con v It includes: starting from the structure layer composed of convolution, normalization, and activation, the middle part has two series-connected Swin-Transformer modules, and the last component is the convolution and SK attention mixed layer CBSK module.

6. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 3 is characterized in that: The multi-scale atrous convolution spatial pyramid module consists of five branches; The first branch uses the residual structure to retain shallow information; The remaining branches first extract features through the dilated convolution layer. Among them, the kernel size of the dilated convolution layer in the four branches is 1×1, and the rest are 3×3, with expansion rates of 1, 6, 12, and 18 respectively. Then, the feature enters a structure consisting of a normalization layer and an activation function in series. Finally, the feature is processed by three consecutive maximum pooling layers and a residual structure. The outputs of these five branches are concatenated to obtain the final feature.

7. The infrared small target detection method based on Transformer and multi-scale dilated convolution spatial pyramid according to claim 3 is characterized in that: Inputting the infrared small target image set of the detected object into the trained infrared small target detection network model based on Transformer and multi-scale dilated convolution spatial pyramid to realize the small target detection of the detected object includes: Use size n ch ×h×w infrared small target image is taken as input into the backbone network, and 5 feature maps of different scales are obtained and output through the convolution layer, convolution and SK attention mixed layer CBSK, movable window transformer and convolution mixed network layer Swin-Conv and multi-scale void convolution spatial pyramid layer. and The scales are and Among them, the number of all Swin-Conv layers is set to 1; The five feature maps of different scales and The neck network is input, and after upsampling, concatenation, C2f layer, and convolution layer processing, 5 feature maps of different scales D1, D2, D3, D4, and D5 are obtained, and the scales are and The five feature maps D1, D2, D3, D4 and D5 of different scales are input into the feature detection head, and five detection heads of different scales are designed in the feature detection head to realize the detection of large targets and small targets.

8. An infrared small target detection system based on Transformer and multi-scale dilated convolutional spatial pyramid, characterized in that: include: Acquisition module, construction module and detection module; The acquisition module is used to obtain a set of infrared small target images of the inspected object; The building module is used to build an infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid; The detection module is used to input the infrared small target image set of the inspected object into the trained infrared small target detection network model based on Transformer and multi-scale hole convolution spatial pyramid to realize small target detection of the inspected object.

Images