Cluster-guided infrared small target hierarchical structure texture fusion segmentation method

Abstract

The application discloses a clustering-guided infrared small target hierarchical structure texture fusion segmentation method, which comprises the following steps: constructing an infrared small target segmentation data set; constructing an initial infrared small target segmentation network model; inputting a training set into the model for training to obtain an infrared small target segmentation network model; inputting a test set or an infrared small target image to be segmented into the infrared small target segmentation network model to output an infrared small target segmentation result, and the like. The application provides a small target segmentation deep learning method for an infrared small target image, which combines image clustering results to perform a composition strategy, integrates image prior structure information in an initial composition stage, effectively extracts infrared image structure information through a graph feature learning module, learns a global-local hierarchical structure relationship of the image, effectively fuses infrared image structure information and texture information through feature pyramid fusion, deepens overall scene cognition of the infrared image, and makes the algorithm more robust.

Description

Technical Field

[0001] This invention belongs to the field of infrared small target segmentation technology, and specifically relates to a cluster-guided hierarchical structure texture fusion segmentation method for infrared small targets. Background Technology

[0002] Any object with a temperature above absolute zero will radiate electromagnetic waves in the infrared band. Radiation in the infrared band cannot be detected by the human visual system, therefore, an infrared detection system is needed for visual imaging. Compared with visible light imaging detection and active radar imaging detection, infrared imaging detection has several advantages: (1) Infrared imaging is not affected by light intensity and can work normally in low-light environments; (2) Infrared detection technology is a passive detection technology, offering stronger concealment; (3) Infrared imaging detection has strong penetration capability, good anti-interference ability, and is not affected by radar stealth and other camouflage technologies. Currently, infrared imaging systems have achieved successful and widespread applications in civilian fields such as medical diagnostic analysis, agricultural and industrial monitoring, and facial recognition. Furthermore, they are widely used in military fields such as military reconnaissance, early warning, and guidance. Infrared small target segmentation is one of the components of this system.

[0003] However, infrared small target segmentation faces numerous difficulties and challenges: ① Infrared small targets are generally very small, ranging from one pixel to tens of pixels in an image; ② Infrared small targets typically have a low signal-to-noise ratio, easily getting lost in high noise and complex backgrounds; ③ Infrared small targets lack shape features, and their size and shape vary greatly in different scenes. Traditional infrared small target segmentation methods use fixed hyperparameter settings, making these methods lack robustness; while existing deep learning-based methods focus on mining color and texture information in infrared images, lacking the extraction of structural information, resulting in unsatisfactory performance in infrared small target images lacking color and texture information. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a cluster-guided method for hierarchical structure texture fusion and segmentation of small infrared targets.

[0005] To achieve the above objectives, the clustering-guided hierarchical structure texture fusion and segmentation method for small infrared targets provided by the present invention includes the following steps performed in sequence:

[0006] S1. Construct an infrared small target segmentation dataset using publicly available or collected infrared small target images, and divide it into a training set and a test set according to the proportions; then perform size unification, left and right flipping, multi-scale transformation and normalization on the infrared small target segmentation images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation.

[0007] S2. Construct an initial infrared small target segmentation network model using the PyTorch deep learning framework. The initial infrared small target segmentation network model includes a clustering segmentation module, a convolutional neural network ResNet, a graph structure generation module, two serially connected graph feature learning modules, a feature pyramid fusion module, and a prediction output head.

[0008] S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target segmentation network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, number of training rounds and number of batches, use Soft-IoU as the loss function, and use the Kaiming method to initialize the weights and biases in the infrared small target segmentation network model to obtain the trained infrared small target segmentation network model.

[0009] S4. After adjusting the size of the test set or the infrared small target image to be segmented obtained in S1 to 256×256 pixels, input it into the trained infrared small target segmentation network model above, and output the final infrared small target segmentation result.

[0010] In S1, the method of constructing an infrared small target segmentation dataset using publicly available or acquired infrared small target images and dividing it into a training set and a test set proportionally; then, the infrared small target segmentation images in the training set are subjected to size unification, left-right flipping, multi-scale transformation, and normalization processing to expand the training set and obtain a data-enhanced infrared small target image training set is as follows:

[0011] Obtain N infrared small target images from a publicly available dataset or original infrared small target images collected. Manually segment and label the original infrared small targets to obtain true labels for the infrared small target images. The infrared small target segmentation dataset is composed of all labeled segmented infrared small target images and divided into training and testing sets in an 8:2 ratio. Then, the labeled segmented infrared small target images in the training set are standardized to 256×256 pixels and flipped horizontally with a probability of 0.5. Next, scale transformations of 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, and 2.0 times are performed. Finally, normalization is applied to obtain the data-augmented infrared small target image training set.

[0012] In S3, the method of inputting the infrared small target image training set obtained in S1 into the initial infrared small target segmentation network model constructed in S2 for training, using an adaptive gradient descent optimizer, setting the learning rate, number of training epochs and batch size, using Soft-IoU as the loss function, and using the Kaiming method to initialize the weights and biases in the infrared small target segmentation network model to obtain the trained infrared small target segmentation network model is as follows:

[0013] S3.1 First, the clustering segmentation module in the infrared small target segmentation network model is used to initially cluster the infrared small target images in the training set into Q clusters, and the clustering segmentation result S∈{0,1,2,…,Q-1} is obtained.

[0014] Given an infrared small target image I∈R 3×H×W Where H and W are the height and width of the infrared small target image, respectively, the infrared small target image is downsampled by 4 times to obtain the reduced infrared small target image I. ’ ∈R 3×H / 4×W / 4 Then, in the reduced infrared small target image I ’ Q pixels are uniformly selected on the plane as cluster centers, and the features of each cluster center are obtained by averaging the features of the surrounding 256 pixels (16×16 pixel area); then, the infrared small target image I is calculated. ’ The cosine similarity matrix C∈R between the features of each pixel and the features of the Q cluster centers. Q×N1 Where N1 = H / 4 × W / 4; finally, the reduced infrared small target image I ’ Each pixel is assigned to the cluster with the highest similarity to the cosine, thus obtaining the clustering segmentation result S∈{0,1,2,…,Q-1}, with each pixel having a corresponding index label;

[0015] S3.2 Extracting infrared small target images I∈R using a convolutional neural network ResNet. 3×H×W The texture features are used to generate three feature maps F1∈R. C1×128×128 Feature map F2∈R C2×64×64 and feature map F3∈R C3×32×32 ,like Figure 2 As shown, each feature map contains texture information from the infrared image;

[0016] S3.3, The feature map F2∈R obtained in S3.2 above. C2×64×64 Each pixel feature is treated as a graph node feature, and the graph structure generation module maps it to the clustering segmentation result S∈{0,1,2,…,Q-1} obtained in S3.1. Then, the graph node features belonging to the same cluster in the clustering segmentation result S∈{0,1,2,…,Q-1} are connected with edges, and they are regarded as neighbors, thereby generating a graph structure G(V,A) for subsequent graph representation learning, where V is the graph node feature matrix and A is the adjacency matrix.

[0017] S3.4. Using two serially connected graph feature learning modules, update the graph node features of the graph structure G(V,A) obtained in S3.3 according to equation (1), using the ReLU function as the activation function, to obtain the updated graph node feature matrix V'; finally, generate the feature map F. S ∈R C2×64×64 and feature map F I ∈R C2×64×64 ;

[0018] (1);

[0019] Where W is a matrix with learnable parameters, D is the degree matrix of the graph nodes, and σ represents the activation function, which is the ReLU function used here;

[0020] S3.5. Use the feature pyramid fusion module to fuse the above feature maps F1, F2, F3, and F4. S and feature map F I Obtain robust feature map F P ∈R C1×128×128 ;

[0021] Feature map F3, after passing through a 1×1 convolutional layer, is compared with feature map F3, which is downsampled by a factor of 2. I Feature stacking and fusion are performed to obtain feature map F. P1 Feature map F P1 After a 1×1 convolution and a 2x upsampling, the feature map is fused with the feature map F2 to obtain feature map F. P2 Feature map F P2 After 1×1 convolution, it is compared with the feature map F. S Feature stacking and fusion are performed to obtain feature map F. P3 Feature map F P3 After a 1×1 convolution and a 2x upsampling, the feature maps are added together and fused to obtain the robust feature map F1. P ∈R C1×128×128 ;

[0022] S3.6. Use the prediction output head to process the above robust feature map F P ∈R C1×128×128 The number of channels is adjusted by 1×1 convolution, and then the image size is adjusted to the original infrared small target image size by 2x upsampling. Then, the Sigmoid function is used to convert the value of each pixel feature into a probability representation, and the image is binarized according to the threshold to obtain the prediction result.

[0023] When the number of training iterations reaches a set threshold or the value of the loss function reaches a set range, the model parameters are considered to be trained and saved, and finally the trained infrared small target segmentation network model is obtained.

[0024] The probability representation refers to the probability value between 0 and 1 obtained by normalizing the pixel features using the Sigmoid function; the threshold is set to 0.5, and pixels with probability values ​​greater than the threshold are considered small infrared targets, while those with lower probabilities are considered background; therefore, the grayscale value of each pixel in the final prediction result Prediction is represented as follows:

[0025] (2);

[0026] The formula for the Soft-IoU loss function is:

[0027] (3);

[0028] Where n is the number of batches, G is the actual segmentation result, P is the segmentation result of the network, and i represents the i-th sample in a batch.

[0029] The clustering-guided hierarchical structure-texture fusion segmentation method for infrared small targets provided by this invention has the following advantages: ① It provides a deep learning method for segmenting small targets by fusing multi-level structural and texture information in infrared small target images; ② The strategy of combining image clustering results for graph construction can incorporate prior structural information of the image in the initial graph construction stage; ③ The designed graph feature learning module can effectively extract structural information of infrared images and learn the global-local hierarchical structure relationship of the image; ④ The progressive feature pyramid fusion can effectively fuse structural and texture information of infrared images, deepen the understanding of the overall scene of infrared images, and make the algorithm more robust. Attached Figure Description

[0030] Figure 1 The flowchart of the clustering-guided hierarchical structure texture fusion and segmentation method for small infrared targets provided by the present invention is shown.

[0031] Figure 2 The diagram shows the structure of a cluster-guided hierarchical structure texture fusion and segmentation method for infrared small targets. Detailed Implementation

[0032] The clustering-guided hierarchical structure texture fusion and segmentation method for infrared small targets provided by the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, the clustering-guided hierarchical structure texture fusion and segmentation method for small infrared targets provided by this invention includes the following steps performed in sequence:

[0034] S1. Construct an infrared small target segmentation dataset using publicly available or collected infrared small target images, and divide it into a training set and a test set according to the proportions; then perform size unification, left and right flipping, multi-scale transformation and normalization on the infrared small target segmentation images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation.

[0035] Obtain N infrared small target images from a publicly available dataset or original infrared small target images collected. Manually segment and label the original infrared small targets to obtain true labels for the infrared small target images. The infrared small target segmentation dataset is composed of all labeled segmented infrared small target images and divided into training and testing sets in an 8:2 ratio. Then, the labeled segmented infrared small target images in the training set are standardized to 256×256 pixels and flipped horizontally with a probability of 0.5. Next, scale transformations of 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, and 2.0 times are performed. Finally, normalization is applied to obtain the data-augmented infrared small target image training set.

[0036] In this invention, the disclosed dataset uses NUAA-SIRST, NUDT-SIRST, and ground / aerial background datasets; N>1000; the infrared small target segmentation dataset SCISTD1 consists of all labeled infrared small target segmentation images; the infrared small target segmentation dataset SCISTD1 contains 3627 infrared small target segmentation images, of which 2901 constitute the training set and the remaining 726 are used as the test set. This dataset includes infrared small target images in various complex scenes such as sky, ground, clouds, and forests, and the size of the infrared small targets is less than 100 pixels.

[0037] S2. Construct an initial infrared small target segmentation network model using the PyTorch deep learning framework. The initial infrared small target segmentation network model includes a clustering segmentation module, a convolutional neural network ResNet, a graph structure generation module, two serially connected graph feature learning modules, a feature pyramid fusion module, and a prediction output head.

[0038] S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target segmentation network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, number of training rounds and number of batches, use Soft-IoU as the loss function, and use the Kaiming method to initialize the weights and biases in the infrared small target segmentation network model to obtain the trained infrared small target segmentation network model.

[0039] The specific method is as follows:

[0040] S3.1 In this invention, the learning rate is set to 0.03, the number of training rounds is 200, and the number of batches is 16. First, the clustering segmentation module in the infrared small target segmentation network model is used to initially cluster the infrared small target images in the input infrared small target image training set into Q clusters, and the clustering segmentation result S∈{0,1,2,…,Q-1} is obtained.

[0041] Given an infrared small target image I∈R 3×H×W Where H and W are the height and width of the infrared small target image, respectively, the infrared small target image is downsampled by 4 times to obtain the reduced infrared small target image I. ’ ∈R 3×H / 4×W / 4 Then, in the reduced infrared small target image I ’ Q pixels are uniformly selected on the plane as cluster centers, and the features of each cluster center are obtained by averaging the features of the surrounding 256 pixels (16×16 pixel area); then, the infrared small target image I is calculated. ’ The cosine similarity matrix C∈R between the features of each pixel and the features of the Q cluster centers. Q×N1 Where N1 = H / 4 × W / 4; finally, the reduced infrared small target image I ’ Each pixel is assigned to the cluster with the highest similarity to the cosine, thus obtaining the clustering segmentation result S∈{0,1,2,…,Q-1}, with each pixel having a corresponding index label;

[0042] S3.2 Extracting infrared small target images I∈R using a convolutional neural network ResNet. 3×H×W The texture features are used to generate three feature maps F1∈R. C1×128×128 Feature map F2∈R C2×64×64 and feature map F3∈R C3×32×32 ,like Figure 2 As shown, each feature map contains texture information from the infrared image;

[0043] S3.3, The feature map F2∈R obtained in S3.2 above. C2×64×64 Each pixel feature is treated as a graph node feature, and the graph structure generation module maps it to the clustering segmentation result S∈{0,1,2,…,Q-1} obtained in S3.1. Then, the graph node features belonging to the same cluster in the clustering segmentation result S∈{0,1,2,…,Q-1} are connected with edges, and they are regarded as neighbors, thereby generating a graph structure G(V,A) for subsequent graph representation learning, where V is the graph node feature matrix and A is the adjacency matrix.

[0044] S3.4. Using two serially connected graph feature learning modules, the graph structure G(V,A) obtained in S3.3 is updated with graph node features according to Equation (1). The ReLU function is used as the activation function to obtain the updated graph node feature matrix V'. The graph node feature matrix V interacts with the features according to the explicit guidance of the adjacency matrix A, which can maintain the semantic consistency of graph node features within a cluster and enhance the discriminability between features of different cluster nodes. Finally, the feature map F is generated. S ∈R C2×64×64 and feature map F I ∈R C2×64×64 ;

[0045] (1);

[0046] Where W is a matrix with learnable parameters, D is the degree matrix of the graph nodes, and σ represents the activation function, which is the ReLU function used here;

[0047] S3.5. Use the feature pyramid fusion module to fuse the above feature maps F1, F2, F3, and F4. S and feature map F I Obtain robust feature map F P ∈R C1×128×128 ;

[0048] like Figure 2 As shown, after extracting feature maps F1, F2, F3, and F4... S and feature map F I Then, the feature pyramid fusion module is used to fuse these feature maps. Specifically, feature map F3 is processed through a 1×1 convolutional layer and then fused with feature map F3 downsampled by a factor of 2. I Feature concat fusion is performed to obtain feature map F. P1 Feature map F P1 After a 1×1 convolution and a 2x upsampling, the feature map is fused with the feature map F2 to obtain feature map F. P2 Feature map F P2 After 1×1 convolution, it is compared with the feature map F. S Feature concat fusion is performed to obtain feature map F. P3 Feature map F P3 After a 1×1 convolution and a 2x upsampling, the feature maps are added together and fused to obtain the robust feature map F1. P ∈R C1×128×128 ;

[0049] S3.6. Use the prediction output head to process the above robust feature map F P ∈R C1×128×128The number of channels is adjusted by 1×1 convolution, and then upsampled by 2 times to the size of the original infrared small target image. Then, the Sigmoid function is used to convert the value of each pixel feature into a probability representation, and binarization is performed according to the threshold to obtain the prediction result.

[0050] The probability representation refers to the probability value between 0 and 1 obtained by normalizing the pixel features using the Sigmoid function. The threshold is set to 0.5; pixels with probability values ​​greater than the threshold are considered small infrared targets, while those with lower probabilities are considered background. Therefore, the grayscale value of each pixel in the final prediction result is represented as follows:

[0051] (2);

[0052] A soft-IoU loss function is used to calculate the loss between the network's predictions and the ground truth labels, addressing the class imbalance problem between small infrared targets and the background, and making the network pay more attention to small infrared targets. The formula for the soft-IoU loss function is:

[0053] (3);

[0054] Where n is the number of batches, G is the actual segmentation result, P is the segmentation result of the network, and i represents the i-th sample in a batch.

[0055] When the number of training iterations reaches a set threshold or the value of the loss function reaches a set range, the model parameters are considered to be trained and saved. Finally, the trained infrared small target segmentation network model is obtained.

[0056] S4. After adjusting the size of the test set or the infrared small target image to be segmented obtained in S1 to 256×256 pixels, input it into the trained infrared small target segmentation network model above, and output the final infrared small target segmentation result.

[0057] To test the effectiveness and superiority of the method of this invention, the inventors conducted experiments using the test set of the infrared small target segmentation dataset SCISTD1 and compared it with six other infrared small target segmentation methods. Two of the comparison methods were traditional methods (PSTNN and FKRW), and four methods were convolutional neural network-based methods (ACM, UIU-Net, MDvsFA, and AGPC-Net). The comparative experimental results are shown in Table 1, with the best results shown in bold. It can be seen that the method of this invention has the highest Intersection over Union (IoU), normalized Intersection over Union (nIoU), and detection rate P0. d and the lowest false alarm rate F a .

[0058] Table 1. Comparison Experiment Results of SCISTD1 Dataset

[0059]

[0060] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A cluster-guided hierarchical structure texture fusion and segmentation method for small infrared targets, characterized in that: The cluster-guided hierarchical structure texture fusion and segmentation method for infrared small targets includes the following steps performed in sequence: S1. Construct an infrared small target segmentation dataset using publicly available or collected infrared small target images, and divide it into a training set and a test set according to the proportions; then perform size unification, left and right flipping, multi-scale transformation and normalization on the infrared small target segmentation images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation. S2. Construct an initial infrared small target segmentation network model using the PyTorch deep learning framework. The initial infrared small target segmentation network model includes a clustering segmentation module, a convolutional neural network ResNet, a graph structure generation module, two serially connected graph feature learning modules, a feature pyramid fusion module, and a prediction output head. S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target segmentation network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, number of training rounds and number of batches, use Soft-IoU as the loss function, and use the Kaiming method to initialize the weights and biases in the infrared small target segmentation network model to obtain the trained infrared small target segmentation network model. Specific methods: S3.1 First, the clustering segmentation module in the infrared small target segmentation network model is used to initially cluster the infrared small target images in the training set into Q clusters, and the clustering segmentation result S∈{0,1,2,…,Q-1} is obtained. S3.2 Extracting infrared small target images I∈R using a convolutional neural network ResNet. 3×H×W The texture features are used to generate three feature maps F1∈R. C1×128×128 Feature map F2∈R C2×64×64 and feature map F3∈R C3×32×32 ; S3.3, The feature map F2∈R obtained in S3.2 above. C2×64×64 Each pixel feature is treated as a graph node feature, and the graph structure generation module maps it to the clustering segmentation result S∈{0,1,2,…,Q-1} obtained in S3.

1. Then, the graph node features belonging to the same cluster in the clustering segmentation result S∈{0,1,2,…,Q-1} are connected by edges, and they are regarded as neighbors, thereby generating a graph structure G(V,A), where V is the graph node feature matrix and A is the adjacency matrix. S3.

4. Using two serially connected graph feature learning modules, update the graph node features of the graph structure G(V,A) obtained in S3.3, using the ReLU function as the activation function, to obtain the updated graph node feature matrix V'; finally, generate the feature map F. S ∈R C2×64×64 and feature map F I ∈R C2×64×64 ; S3.

5. Use the feature pyramid fusion module to fuse the above feature maps F1, F2, F3, and F4. S and feature map F I Obtain robust feature map F P ∈R C1×128×128 ; S3.

6. Use the prediction output head to process the above robust feature map F P ∈R C1×128×128 The number of channels is adjusted by 1×1 convolution, and then the image size is adjusted to the original infrared small target image size by 2x upsampling. Then, the Sigmoid function is used to convert the value of each pixel feature into a probability representation, and the image is binarized according to the threshold to obtain the prediction result. When the number of training iterations reaches a set threshold or the value of the loss function reaches a set range, the model parameters are considered to be trained and saved. Finally, the trained infrared small target segmentation network model is obtained. S4. After adjusting the size of the test set or the infrared small target image to be segmented obtained in S1 to 256×256 pixels, input it into the trained infrared small target segmentation network model above, and output the final infrared small target segmentation result.

2. The cluster-guided hierarchical structure texture fusion and segmentation method for small infrared targets according to claim 1, characterized in that: In S1, the method of constructing an infrared small target segmentation dataset using publicly available or acquired infrared small target images and dividing it into a training set and a test set proportionally; then, the infrared small target segmentation images in the training set are subjected to size unification, left-right flipping, multi-scale transformation, and normalization processing to expand the training set and obtain a data-enhanced infrared small target image training set is as follows: Obtain N infrared small target images from a publicly available dataset or original infrared small target images collected. Manually segment and label the original infrared small targets to obtain true labels for the infrared small target images. The infrared small target segmentation dataset is composed of all labeled segmented infrared small target images and divided into training and testing sets in an 8:2 ratio. Then, the labeled segmented infrared small target images in the training set are standardized to 256×256 pixels and flipped horizontally with a probability of 0.

5. Next, scale transformations of 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, and 2.0 times are performed. Finally, normalization is applied to obtain the data-augmented infrared small target image training set.

3. The cluster-guided hierarchical structure texture fusion and segmentation method for small infrared targets according to claim 1, characterized in that: In S3.1, the method for initially clustering the infrared small target images in the input infrared small target image training set into Q clusters using the clustering segmentation module in the infrared small target segmentation network model to obtain the clustering segmentation result S∈{0,1,2,…,Q-1} is as follows: Given an infrared small target image I∈R 3×H×W Where H and W are the height and width of the infrared small target image, respectively, the infrared small target image is downsampled by 4 times to obtain the reduced infrared small target image I. ’ ∈R 3×H / 4×W / 4 Then, in the reduced infrared small target image I ’ Q pixels are uniformly selected on the plane as cluster centers, and the features of each cluster center are obtained by averaging the features of its surrounding 256 pixels; then, the infrared small target image I is calculated. ’ The cosine similarity matrix C∈R between the features of each pixel and the features of the Q cluster centers. Q×N1 Where N1 = H / 4 × W / 4; finally, the reduced infrared small target image I ’ Each pixel is assigned to the cluster with the highest similarity to the cosine, thus obtaining the clustering segmentation result S∈{0,1,2,…,Q-1}, with each pixel having a corresponding index label; In S3.4, the formula for calculating the updated graph node feature matrix V' is: （1）； Where W is a matrix with learnable parameters, D is the degree matrix of the graph nodes, and σ represents the activation function; In S3.5, the feature pyramid fusion module is used to fuse the aforementioned feature maps F1, F2, F3, and F4. S and feature map F I Obtain robust feature map F P ∈R C1×128×128 The method is: Feature map F3, after passing through a 1×1 convolutional layer, is compared with feature map F3, which is downsampled by a factor of 2. I Feature stacking and fusion are performed to obtain feature map F. P1 Feature map F P1 After a 1×1 convolution and a 2x upsampling, the feature map is fused with the feature map F2 to obtain feature map F. P2 Feature map F P2 After 1×1 convolution, it is compared with the feature map F. S Feature stacking and fusion are performed to obtain feature map F. P3 Feature map F P3 After a 1×1 convolution and a 2x upsampling, the feature maps are added together and fused to obtain the robust feature map F1. P ∈R C1×128×128 .

4. The cluster-guided hierarchical structure texture fusion and segmentation method for small infrared targets according to claim 3, characterized in that: The probability representation refers to the probability value between 0 and 1 obtained by normalizing the pixel features using the Sigmoid function; the threshold is set to 0.5, and pixels with probability values ​​greater than the threshold are considered infrared small targets, while those with lower probabilities are considered background; therefore, the grayscale value of each pixel in the final prediction result Prediction is represented as follows: (2)； The formula for the Soft-IoU loss function is: (3)； Where n is the number of batches, G is the actual segmentation result, P is the segmentation result of the network, and i represents the i-th sample in a batch.

Images