Double-coding fusion infrared small target detection method based on dynamic convolution and Mama

Abstract

The invention discloses an infrared small target detection method based on double-coding fusion of dynamic convolution and Mamba. The method comprises the following steps: constructing an initial infrared small target detection network model; inputting the training set into a model for training to obtain an infrared small target detection network model; and inputting the infrared small target image of the test set into the infrared small target detection network model, and outputting an infrared small target detection result. The method comprises the following steps: extracting texture structure information of an infrared small target by using a local information encoder constructed by a dynamic convolution method; a global information encoder constructed by using a Mama method is used for modeling large-view range information of an image to obtain effective global context information; the local and global feature information of the infrared image can be effectively fused through up-sampling cross enhancement fusion, the correlation of the local and global features is enhanced, the semantic difference between the target and the background clutter is enhanced, and the algorithm detection performance and robustness are further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared small target detection, and particularly relates to an infrared small target detection method based on dual coding fusion of dynamic convolution and Mamba. Background Art

[0002] Any object with a temperature higher than absolute zero will radiate electromagnetic waves in the infrared band. The radiation light in the infrared band cannot be sensed by the human visual system, so an infrared detection system needs to be used for visual imaging. Compared with visible light imaging detection and active radar imaging detection, infrared imaging detection has multiple advantages: (1) The infrared detection system does not require an external light source, can image at night and in low light conditions, has strong penetration, and is suitable for target detection and monitoring in bad weather; (2) Infrared detection technology belongs to passive detection technology and has stronger concealment; (3) Infrared thermal imaging technology uses thermal infrared rays, so it is not affected by electromagnetic interference, can accurately track thermal targets at a long distance, and can accurately guide. Currently, infrared imaging systems are widely used in fields such as maritime rescue, ground monitoring, and flight navigation. And infrared small target detection is one of the components.

[0003] However, due to the influence of atmospheric interference and long-distance detection, infrared small target detection faces many difficulties and challenges: ① Infrared small targets often only occupy a small part of the infrared image, and the area occupied in the image ranges from one pixel to dozens of pixels; ② Factors such as atmospheric interference and radiation energy attenuation result in low target gray values, and it is easy to be submerged by background noise; ③ Infrared small targets lack significant shape features, and their sizes and shapes vary greatly in different scenarios. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide an infrared small target detection method based on dual coding fusion of dynamic convolution and Mamba.

[0005] In order to achieve the above purpose, the infrared small target detection method based on dual coding fusion of dynamic convolution and Mamba provided by the present invention includes the following steps carried out in sequence:

[0006] S1. After annotating the publicly available infrared small target images, construct an infrared small target detection data set, and divide it into a training set and a test set according to a ratio; then perform size unification, up-down, left-right flipping, and normalization processing on the infrared small target images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation;

[0007] S2. Use the Pytorch deep learning framework to construct an initial infrared small target detection network model, and the initial infrared small target detection network model includes a dynamic convolution encoder, a vision Mamba encoder, and a decoder;

[0008] S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target detection network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, number of training epochs, and batch size, adopt Soft-IoU as the loss function, and use the kaiming method to initialize the weights and bias terms in the above initial infrared small target detection network model to obtain a trained infrared small target detection network model;

[0009] S4. Resize the infrared small target images in the test set obtained in S1 to a size of 256×256 pixels, and then input them into the above trained infrared small target detection network model to output the final infrared small target detection results.

[0010] In S1, the method of annotating the publicly available infrared small target images to construct an infrared small target detection data set and dividing it into a training set and a test set according to a ratio; then unifying the sizes, flipping left and right, performing multi-scale transformation, and normalizing the infrared small target images in the training set to expand the training set and obtain the infrared small target image training set after data augmentation is as follows:

[0011] Obtain N infrared small target images from the publicly available data set, use manual annotation to obtain the true labels of the infrared small targets in them, and form an infrared small target detection data set from all the labeled infrared small target images, and divide it into a training set and a test set according to a ratio of 8:2; then unify the sizes of the labeled infrared small target images in the training set to 256×256 pixels, then perform up, down, left, and right flips with a flipping probability of 0.6, and finally perform normalization processing to obtain the infrared small target image training set after data augmentation.

[0012] In S2, the dynamic convolutional encoder is composed of a series of dynamic convolutional modules, and each dynamic convolutional module is stacked by a conditional convolutional layer, a batch normalization layer, and an activation function layer; the visual Mamba encoder is composed of a feature embedding module, a series of PatchMerging layers, and a state space module, where the state space module includes a normalization layer, a VSSBlock module, a conditional convolutional module, and a linear transformation layer; the decoder is composed of multiple upsampling fusion modules, and each upsampling fusion module includes a conditional convolutional module, an upsampling module, and a linear transformation layer.

[0013] In S3, the method of inputting the infrared small target image training set obtained in S1 into the initial infrared small target detection network model constructed in S2 for training, using the adaptive gradient descent optimizer, setting the learning rate, the number of training epochs, and the batch size, adopting Soft-IoU as the loss function, and initializing the weights and bias terms in the above initial infrared small target detection network model using the kaiming method to obtain the trained infrared small target detection network model is as follows:

[0014] S3.1. First, the infrared small target image I ∈ R after data augmentation is input into the conditional convolution layer of the dynamic convolution module in the dynamic convolution encoder for global average pooling to obtain global statistical information, and the result of global average pooling is mapped through the batch normalization layer to obtain the weights corresponding to n experts. Then, the activation function softmax in the activation function layer is used to normalize the weights so that the sum of the weights is 1. Then, the torch.einsum function is used to multiply the obtained weights of the n experts by the weights initialized using the kaiming method to obtain the dynamic convolution kernel related to the input infrared small target image after data augmentation. Next, the dynamic convolution kernel is used to perform a convolution operation on the infrared small target image after data augmentation, and a series of dimensionality transformations are performed to obtain local features 3×H×W The local features Local features And local features where c represents the dimension of the local features obtained after dimensionality transformation, H represents the height of the infrared small target image after data augmentation, W represents the width of the infrared small target image after data augmentation, and each feature contains the local texture features of the infrared small target image;

[0015] S3.2. The infrared small target image I ∈ R after data augmentation is 3×H×WInput the Visual Mamba encoder. Part of it is directly transmitted, and the other part is normalized through a normalization layer (LN) to obtain normalized features. Then, the normalized features are respectively input into two parallel branches, the conditional convolution module and the VSSBlock module, for processing to obtain the global features of the infrared small target image. The normalized features are processed by the conditional convolution module and the linear transformation layer to obtain the output features of the conditional convolution module branch. At the same time, the normalized features enter the VSSBlock module branch. First, the split function is used to split the channel dimension to obtain two features, feature 1 and feature 2. Among them, feature 1 is processed by the linear transformation layer, the depthwise separable convolution layer, the activation function layer, and the SS2D module to obtain output feature 1. Feature 2 is processed by the linear layer and uses the dynamic activation function APA as the activation function to obtain output feature 2. Then, each element of output feature 1 and output feature 2 is multiplied element by element through dot multiplication to obtain output feature 3. Then, output feature 3 is linearly transformed by the linear layer to obtain the output features of the VSSBlock module branch. Finally, the directly transmitted features, the output features of the conditional convolution module, and the output features of the VSSBlock module are fused in a summation manner to obtain the global features;

[0016] S3.3. Input the local features and global features obtained in the above S4.1 and S4.2 into the upsampling fusion module of the corresponding layer of the decoder. First, use the conditional convolution module to extract high-level global semantic features from the global features and low-level local spatial features from the local features. Then, use the dot multiplication method to achieve the interaction enhancement between the local features and the global features to obtain the interactively enhanced features. Then, use the residual connection method to add the input local features and global features to the interactively enhanced features respectively to obtain the enhanced local features and global features. Next, since the decoder has multiple upsampling fusion modules, when the first upsampling fusion module starts upsampling, there is no fusion feature from the previous layer, and only the concat method is used to upsample the enhanced local features and global features. Starting from the second upsampling fusion module, when upsampling, the fusion feature from the previous layer is concatenated with the enhanced local features and global features of the current layer in the channel dimension using the concat method. Finally, a linear layer is used for dimension adjustment to obtain the final fusion feature;

[0017] S3.4. After adjusting the number of channels of the final fusion feature output by the last upsampling fusion module using a 1×1 convolution, use the Sigmoid function to convert the value of each pixel point feature into a probability representation, and perform binarization according to the threshold to obtain the final prediction result;

[0018] When the number of training times reaches the set threshold or the value of the loss function reaches the set range, it is considered that the model parameters are trained, and the model parameters are saved, and finally a trained infrared small target detection network model is obtained.

[0019] In S3.4, the probability representation is the output value between 0 and 1 obtained by normalizing the pixel point feature after passing through the Sigmoid function; the threshold is set to 0.5, and the pixel points with gray values greater than the threshold are regarded as infrared small targets, and vice versa as the background; therefore, the gray value of each pixel point in the final prediction result Prediction is expressed as:

[0020]

[0021] In S3, the formula of the Soft-IoU loss function is:

[0022]

[0023] Among them, n is the number of batch processing, G is the true label, P is the result predicted by the network, and i represents the i-th sample in a batch.

[0024] The infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba provided by the present invention has the following advantages: ① A deep learning method for fusing local feature information and global feature information to achieve infrared small target detection is given for infrared small target images; ② An encoder constructed using the dynamic convolution method is designed to extract the texture structure information of infrared small targets, which can overcome the problem of the fixed receptive field of a single convolution kernel in the traditional convolution method and fully obtain the detailed texture information of infrared small targets; ③ A global information encoder constructed using the Mamba method is designed to obtain effective global context information by modeling the information in the large field of view of the image; ④ The upsampling fusion module uses the low-level local spatial features and high-level global semantic features of infrared small targets through the method of cross-enhancement, enhances the correlation between local and global features, and at the same time strengthens the semantic difference between the target and background clutter, thereby improving the detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of the infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba provided by the present invention.

[0026] Figure 2 It is a structural diagram of the initial infrared small target detection network model in the present invention.

[0027] Figure 3 It is a structural diagram of the dynamic convolution encoder in the present invention.

[0028] Figure 4 It is a structural diagram of the state space module in the present invention.

[0029] Figure 5 This is the structural diagram of the upsampling fusion module in the present invention. Detailed implementation manners

[0030] The following combines the accompanying drawings and specific embodiments to detail the infrared small target detection method based on dual-coding fusion of dynamic convolution and Mamba provided by the present invention.

[0031] As Figure 1 shown, the infrared small target detection method based on dual-coding fusion of dynamic convolution and Mamba provided by the present invention includes the following steps carried out in sequence:

[0032] S1. After annotating the publicly available infrared small target images, construct an infrared small target detection data set, and divide it into a training set and a test set according to a ratio; then unify the sizes, flip up, down, left, and right, and normalize the infrared small target images in the training set to expand the training set and obtain the infrared small target image training set after data augmentation;

[0033] Obtain N infrared small target images from the publicly available data set, use manual annotation of the infrared small targets therein to obtain the true labels of the infrared small target images, and form an infrared small target detection data set from all the labeled infrared small target images, and divide it into a training set and a test set according to a ratio of 8:2; then unify the sizes of the labeled infrared small target images in the training set to 256×256 pixels, then flip up, down, left, and right with a probability of 0.6, and finally perform normalization processing to obtain the infrared small target image training set after data augmentation.

[0034] In the present invention, the publicly available data sets adopt the NUAA-SIRST data set and the IRSDT-1k data set; each data set has infrared images of short-wave wavelength, medium-wave wavelength, and 950nm wavelength, and the data sets include scenes such as sky, ocean, and buildings; the NUAA-SIRST data set includes 427 representative images from various real scenes; the IRSDT-1k data set includes 1001 manually labeled real-world infrared scene images with different target shapes, sizes, and cluttered backgrounds from different scenes; both data sets are divided into a training set and a test set according to a ratio of 8:2.

[0035] S2. Use the Pytorch deep learning framework to construct an initial infrared small target detection network model as Figure 2 shown, the initial infrared small target detection network model includes a dynamic convolution encoder, a Vision Mamba encoder, and a decoder;

[0036] The dynamic convolution encoder is composed of a series of dynamic convolution modules, such asFigure 3 As shown in the figure, each dynamic convolution module is stacked by a conditional convolution layer, a batch normalization layer, and an activation function layer; the Vision Mamba encoder is composed of a feature embedding module, a series of PatchMerging layers, and a state space module, as Figure 4 shown in the figure, where the state space module includes a normalization layer, a VSSBlock module, a conditional convolution module, and a linear transformation layer; the decoder is composed of multiple upsampling fusion modules, as Figure 5 shown in the figure, and each upsampling fusion module includes a conditional convolution module, an upsampling module, and a linear transformation layer.

[0037] S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target detection network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, the number of training epochs, and the batch size, adopt Soft-IoU as the loss function, and use the kaiming method to initialize the weights and bias terms in the above initial infrared small target detection network model to obtain a trained infrared small target detection network model;

[0038] The specific method is as follows:

[0039] S3.1. In the present invention, the learning rate is set to 0.01, the number of training epochs is 200, and the batch size is 8; first, the infrared small target image I ∈ R 3×H×W after data augmentation is input into the conditional convolution layer of the dynamic convolution module in the dynamic convolution encoder for global average pooling to obtain global statistical information, and the result of global average pooling is mapped through the batch normalization layer to obtain the weights corresponding to n experts. Then, the activation function softmax in the activation function layer is used to normalize the weights so that the sum of the weights is 1; then, the torch.einsum function is used to multiply the n experts' weights obtained with the weights initialized by the kaiming method to obtain a dynamic convolution kernel related to the input infrared small target image after data augmentation; then, the dynamic convolution kernel is used to perform a convolution operation on the infrared small target image after data augmentation, and after a series of dimensionality transformations, local features as Figure 2 shown in the figure are obtained Local features and local features are obtained. Among them, c represents the dimension of the local features obtained after dimensionality transformation, H represents the height of the infrared small target image after data augmentation, W represents the width of the infrared small target image after data augmentation, and each feature contains the local texture features of the infrared small target image;

[0040] S3.2. As Figure 4 shown in the figure, the infrared small target image I ∈ R 3×H×WInput the Visual Mamba encoder. Part of it is directly transmitted, and the other part is normalized through a normalization layer (LN) to obtain normalized features. Then, the normalized features are respectively input into two parallel branches of a conditional convolution module and a VSSBlock module for processing to obtain the global features of the infrared small target image. The normalized features are processed by the conditional convolution module and a linear transformation layer to obtain the output features of the conditional convolution module branch. At the same time, the normalized features enter the VSSBlock module branch. First, the split function is used to split the channel dimension to obtain two features, namely feature 1 and feature 2. Among them, feature 1 is processed by a linear transformation layer, a depthwise separable convolution layer, an activation function layer, and an SS2D module to obtain output feature 1. Feature 2 is processed by a linear layer and using the dynamic activation function APA as the activation function to obtain output feature 2. Then, each element of output feature 1 and output feature 2 is multiplied element by element through dot multiplication to obtain output feature 3. Then, output feature 3 is linearly transformed by a linear layer to obtain the output features of the VSSBlock module branch. Finally, the directly transmitted features, the output features of the conditional convolution module, and the output features of the VSSBlock module are fused by summation to obtain the global features;

[0041] S3.3. Input the local features and global features obtained in the above S4.1 and S4.2 into the upsampling fusion module of the corresponding layer of the decoder. As Figure 5 shown, first use the conditional convolution module to extract high-level global semantic features from the global features and low-level local spatial features from the local features. Then, use the dot multiplication method to achieve the interactive enhancement of the local features and global features to obtain the interactively enhanced features. Then, use the residual connection method to add the input local features and global features to the interactively enhanced features respectively to obtain the enhanced local features and global features. Next, since the decoder has multiple upsampling fusion modules, when the first upsampling fusion module starts upsampling, there is no fusion feature from the previous layer, and only the concat method is used to upsample the enhanced local features and global features. Starting from the second upsampling fusion module, when upsampling, the fusion feature from the previous layer is concatenated with the enhanced local features and global features of the current layer in the channel dimension using the concat method. Finally, a linear layer is used for dimension adjustment to obtain the final fusion feature;

[0042] S3.4. After adjusting the number of channels of the final fusion feature output by the last upsampling fusion module using a 1×1 convolution, use the Sigmoid function to convert the value of each pixel point feature into a probability representation, and perform binarization according to the threshold to obtain the final prediction result;

[0043] The probability representation is the output value between 0 and 1 obtained by normalizing the pixel point features after passing through the Sigmoid function; the threshold is set to 0.5, and the pixel points with gray values greater than the threshold are regarded as infrared small targets, and vice versa as the background; therefore, the gray value of each pixel point in the final prediction result Prediction is expressed as:

[0044]

[0045] The Soft-IoU loss function is used to calculate the loss between the network prediction and the ground truth label to handle the class imbalance problem between infrared small targets and the background and make the network pay more attention to infrared small targets; the formula of the Soft-IoU loss function is:

[0046]

[0047] where n is the number of batches, G is the ground truth label, P is the prediction result of the network, and i represents the i-th sample in a batch.

[0048] When the number of training times reaches the set threshold or the value of the loss function reaches the set range, it is considered that the model parameters are trained, the model parameters are saved, and finally the trained infrared small target detection network model is obtained.

[0049] S4. Resize the infrared small target images in the test set obtained in S1 to a size of 256×256 pixels, and then input them into the above-trained infrared small target detection network model to output the final infrared small target detection result.

[0050] To test the effectiveness and superiority of the method of the present invention, the inventor of the present invention uses the test sets of the publicly available infrared small target datasets NUAA-SIRST and IRSTD-1k for experimental tests and compares them with 6 other infrared small target detection methods. Among them, 2 comparison methods are traditional methods (PSTNN, FKRW), and 8 methods are deep learning-based methods (ACM, UIU-Net, AGPC-Net, ALC-Net, DNANet, MAD-Net, IAANet, and IRGraphSeg). The comparison experimental results are shown in Table 1, and the best results are shown in bold. It can be seen that the method of the present invention has the highest indicators in the intersection over union IoU, normalized intersection over union nIoU, and detection rate P d , reaching the highest index, and has the second-best effect in the false alarm rate F a .

[0051] Table 1 Quantitative results of different methods on the NUAA-SIRST and IRSTD-1k datasets

[0052]

[0053]

[0054] In addition to the above embodiments, the present invention may also have other implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope claimed by the present invention.

Claims

1. An infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba, characterized in that: The infrared small target detection method based on dual coding fusion of dynamic convolution and Mamba includes the following steps carried out in sequence: S1. Annotate the publicly available infrared small target images to construct an infrared small target detection dataset, and divide it into a training set and a test set according to a ratio; then unify the sizes, flip horizontally and vertically, and normalize the infrared small target images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation; S2. Use the Pytorch deep learning framework to construct an initial infrared small target detection network model, where the initial infrared small target detection network model includes a dynamic convolution encoder, a vision Mamba encoder, and a decoder; S3. Input the infrared small target image training set obtained in S1 into the initial infrared small target detection network model constructed in S2 for training. Use the adaptive gradient descent optimizer, set the learning rate, the number of training epochs, and the batch size, adopt Soft-IoU as the loss function, and use the kaiming method to initialize the weights and bias terms in the above initial infrared small target detection network model to obtain a trained infrared small target detection network model; S4. Resize the infrared small target images in the test set obtained in S1 to a size of 256×256 pixels, and then input them into the above trained infrared small target detection network model to output the final infrared small target detection results.

2. The infrared small target detection method based on dual-coding fusion of dynamic convolution and Mamba according to claim 1, characterized in that: In S1, the method of annotating the publicly available infrared small target images to construct an infrared small target detection dataset, dividing it into a training set and a test set according to a ratio; then unifying the sizes, flipping horizontally, performing multi-scale transformation, and normalizing the infrared small target images in the training set to expand the training set and obtain an infrared small target image training set after data augmentation is as follows: Obtain N infrared small target images from the publicly available dataset, manually annotate the infrared small targets among them to obtain the true labels of the infrared small target images, and form an infrared small target detection dataset from all the labeled infrared small target images, and divide it into a training set and a test set according to a ratio of 8:2; then unify the sizes of the labeled infrared small target images in the training set to 256×256 pixels, then flip horizontally and vertically with a probability of 0.6, and finally perform normalization to obtain an infrared small target image training set after data augmentation.

3. The infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba according to claim 1, characterized in that: In S2, the dynamic convolution encoder is composed of a series of dynamic convolution modules, and each dynamic convolution module is stacked by a conditional convolution layer, a batch normalization layer, and an activation function layer; the vision Mamba encoder is composed of a feature embedding module, a series of PatchMerging layers, and state space modules, where the state space module includes a normalization layer, a VSSBlock module, a conditional convolution module, and a linear transformation layer; the decoder is composed of multiple upsampling fusion modules, and each upsampling fusion module includes a conditional convolution module, an upsampling module, and a linear transformation layer.

4. The infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba according to claim 1, wherein: In S3, the method of inputting the infrared small target image training set obtained in S1 into the initial infrared small target detection network model constructed in S2 for training, using the adaptive gradient descent optimizer, setting the learning rate, the number of training epochs, and the batch size, adopting Soft-IoU as the loss function, and initializing the weights and bias terms in the above initial infrared small target detection network model using the kaiming method to obtain the trained infrared small target detection network model is as follows: S3.

1. First, the infrared small target image I ∈ R after data augmentation is 3×H×W input into the conditional convolution layer of the dynamic convolution module in the dynamic convolution encoder for global average pooling to obtain global statistical information, and the result of global average pooling is mapped through the batch normalization layer to obtain the weights corresponding to n experts. Then, the activation function softmax in the activation function layer is used to normalize the weights so that the sum of the weights is 1. Then, the torch.einsum function is used to multiply the weights of the n experts obtained with the weights initialized using the kaiming method to obtain a dynamic convolution kernel related to the input infrared small target image after data augmentation. Next, the dynamic convolution kernel is used to perform a convolution operation on the infrared small target image after data augmentation, and local features are obtained through a series of dimensionality transformations Local features and local features where c represents the dimension of the local features obtained after dimensionality transformation, H represents the height of the infrared small target image after data augmentation, W represents the width of the infrared small target image after data augmentation, and each feature contains the local texture features of the infrared small target image; S3.

2. Input the infrared small target image I ∈ R after data augmentation 3×H×W into the Vision Mamba encoder. A part of it is directly transmitted, and the other part is normalized through a normalization layer to obtain the normalized features. Then, the normalized features are respectively input into two parallel branches of the conditional convolution module and the VSSBlock module for processing to obtain the global features of the infrared small target image. The normalized features are processed by the conditional convolution module and the linear transformation layer to obtain the output features of the conditional convolution module branch. At the same time, the normalized features enter the VSSBlock module branch. First, the split function is used to split the channel dimension to obtain two features, namely feature 1 and feature 2. Among them, feature 1 is processed by the linear transformation layer, the depthwise separable convolution layer, the activation function layer, and the SS2D module to obtain the output feature 1. Feature 2 is processed by the linear layer and the dynamic activation function APA as the activation function to obtain the output feature 2. Then, each element of the output feature 1 and the output feature 2 is multiplied element by element through dot multiplication to obtain the output feature 3. Then, the output feature 3 is linearly transformed by the linear layer to obtain the output features of the VSSBlock module branch. Finally, the directly transmitted features, the output features of the conditional convolution module, and the output features of the VSSBlock module are fused in a summation manner to obtain the global features; S3.

3. Input the local features and global features obtained in S4.1 and S4.2 into the upsampling fusion module of the corresponding layer of the decoder. First, use the conditional convolution module to extract high-level global semantic features from the global features and low-level local spatial features from the local features. Then, use the method of dot multiplication to achieve the interactive enhancement of the local features and global features, and obtain the interactively enhanced features. Then, use the residual connection method to add the input local features and global features to the interactively enhanced features respectively to obtain the enhanced local features and global features. Next, since the decoder has multiple upsampling fusion modules, when the first upsampling fusion module starts upsampling, there is no fusion feature from the previous layer, and only the concat method is used for the upsampling of the enhanced local features and global features. Starting from the second upsampling fusion module, when upsampling, the fusion feature from the previous layer is concatenated with the enhanced local features and global features of the current layer in the channel dimension using the concat method. Finally, a linear layer is used for dimension adjustment to obtain the final fusion feature; S3.

4. After adjusting the number of channels of the final fusion feature output by the last upsampling fusion module using a 1×1 convolution, use the Sigmoid function to convert the value of each pixel feature into a probability representation, and perform binarization according to the threshold to obtain the final prediction result; When the number of training times reaches the set threshold or the value of the loss function reaches the set range, it is considered that the model parameters are trained, and the model parameters are saved, and finally the trained infrared small target detection network model is obtained.

5. The infrared small target detection method based on dual-coding fusion of dynamic convolution and Mamba according to claim 4, characterized in that: In S3.4, the probability representation is the output value between 0 and 1 obtained by normalizing the pixel feature after passing through the Sigmoid function; the threshold is set to 0.5, and the pixel points with gray values greater than the threshold are regarded as infrared small targets, and vice versa as the background; therefore, the gray value of each pixel point in the final prediction result Prediction is represented as:

6. The infrared small target detection method based on dual-encoding fusion of dynamic convolution and Mamba according to claim 1, characterized in that: In S3, the formula of the Soft-IoU loss function is: where n is the batch size, G is the true label, P is the result predicted by the network, and i represents the i-th sample in a batch.

Images