An infrared small target detection method based on diversified feature learning and coordination

Abstract

This invention proposes an infrared small target detection method based on diversified feature learning and coordination. The method proposes a diversified feature capture and coordination network that learns and coordinates diversified features through multi-path encoding. This network includes an infrared image reconstruction branch that runs in parallel with the detection branch, preserving small target information and reducing feature loss through complementary context encoding. Furthermore, this invention introduces a global feature extraction branch using FTConv to capture target edges and suppress background noise. A cross-layer adaptive feature selection method adaptively coordinates features across layers, enhancing detection capabilities in complex backgrounds while preserving small target details. This invention also proposes a coordinate calibration loss function and a two-stage training strategy to refine the predicted target location. Experimental results on three infrared small target datasets demonstrate that the proposed method outperforms current state-of-the-art methods.

Description

Technical Field

[0001] This invention belongs to the field of infrared target detection technology, and relates to the research on infrared small target detection methods through the coordination of multiple features. Background Technology

[0002] Infrared imaging possesses powerful detection capabilities and high sensitivity to targets under low-light conditions, thus finding wide application in military, agricultural, and civilian fields. The goal of Infrared Small Target Detection (IRSTD) is to extract small targets from infrared images against complex backgrounds. However, IRSTD faces many challenges due to long imaging distances and unfavorable imaging conditions:

[0003] (1) Small target size: Small targets occupy very few pixels in an image, resulting in weak signal strength and easy obscuring by noise. Typically, targets occupy only about 0.05% of the total pixels in an infrared image.

[0004] (2) Complex background: A complex background environment contains many interfering factors, such as temperature changes, ambient noise, and other heat sources that produce infrared radiation similar to the target signal. A complex background makes it difficult to separate the target signal from the noisy background.

[0005] (3) Lack of texture features: Infrared imaging mainly relies on thermal radiation, and infrared targets are usually small, lacking the regular shapes and textures that can be seen under visible light.

[0006] Existing IRSTD methods can be broadly categorized into traditional methods and deep learning-based methods. Traditional methods include filtering-based methods, methods based on human visual system models, and low-rank methods. These traditional methods rely heavily on prior knowledge, empirical assumptions, and manually tuned hyperparameters, which limits their effectiveness in handling complex real-world scenarios.

[0007] In recent years, deep learning methods have demonstrated superior performance in the IRSTD task. For example, Liu et al. proposed the first IRSTD method based on a convolutional neural network (CNN) using a multilayer perceptron (MLP) for detection. Furthermore, generative adversarial networks (GANs) have also been applied to this task, such as MDvsFA-cGAN, which uses adversarial training to balance false negatives and false positives, thereby improving detection performance. Recent research has shown that treating the IRSTD task as a semantic segmentation problem is generally more effective than traditional detection methods, enabling more accurate identification and localization of small targets. For instance, ACMNet was the first IRSTD method based on semantic segmentation, introducing an asymmetric context module to replace skip connections, achieving cross-layer feature fusion and improving detection performance. DNANet employs a densely nested interaction approach to facilitate feature interaction between high-level and low-level features. ISNet supplements the precise shape information of infrared targets by integrating Taylor finite difference heuristic blocks and bidirectional attention aggregation modules. FC3-Net explores cross-layer correlation and feature compensation to address the loss of detail caused by downsampling layers. DMFNet employs a dual encoder to obtain more effective information, thereby enhancing its ability to detect small targets.

[0008] Despite significant progress in the IRSTD domain, the inherent characteristics of infrared images and targets still present considerable challenges. Specifically, targets in infrared images often lack well-defined shapes and textures, making it difficult to effectively extract multi-class target information. This limitation directly impacts the accuracy of detection and localization of small, irregularly shaped targets. Currently, most detection networks, such as U-Net-based models, typically reduce spatial resolution through downsampling. While this helps extract deeper features, it also leads to the loss of local contextual information, hindering the network from recovering and utilizing subtle target details. Although existing network architectures can handle regular targets, they struggle with detail loss, especially when dealing with irregularly shaped and textureless targets. Therefore, advancing network architectures and feature learning strategies, particularly in preserving local details and recovering subtle target features, has become a current research focus. Inspired by the core idea of ​​preserving and recognizing fine target features for reconstruction, we propose an innovative infrared image reconstruction mechanism to capture and retain important information from the original image. Thus, we are the first to apply an infrared image reconstruction mechanism to the IRSTD task. Specifically, we propose a novel network—the Diverse Feature Capture and Coordination Network—to recover and optimize target information, thereby improving the performance of infrared small target detection. Summary of the Invention

[0009] The purpose of this invention is to provide a detection method for infrared small target detection. With the application of deep learning in infrared small target detection, significant progress has been made in detection accuracy. Despite significant advancements in the IRSTD domain, the inherent features of infrared images and targets still present significant challenges. Specifically, targets in infrared images often lack well-defined shapes and textures, making it difficult to effectively extract multi-class target information. This limitation directly affects the accuracy of detection and localization of small, irregularly shaped targets. Currently, most detection networks, such as U-Net-based models, typically reduce spatial resolution through downsampling. While this helps extract deeper features, it also leads to the loss of local contextual information, hindering the network from recovering and utilizing subtle target details. Therefore, this invention proposes a diversified feature capture and coordination network for recovering and optimizing target information to improve the performance of infrared small target detection. Its core consists of the following four parts:

[0010] 1. FFT Hybrid Encoded Block

[0011] Infrared images often suffer from low contrast and complex backgrounds, making it difficult for traditional convolutional methods to extract subtle target features. Convolutional neural networks (CNNs) can also be affected by background noise, impacting the accuracy and performance of Infrared Target Detection (IRSTD). Spectral convolution, by utilizing Fourier transform to convert convolution operations into frequency domain operations, significantly improves feature extraction. The Fourier transform provides additional frequency information, aiding in the identification of targets in low-contrast regions, while also processing background information in the frequency domain, thus separating the target from the background and improving detection accuracy. To improve computational efficiency, we use the Fast Fourier Transform (FFT).

[0012] We designed a hybrid encoding block based on FFT to effectively extract local and global features. The specific architecture of the encoding block is as follows: Figure 1 As shown in (a) and (b), we transform the input feature X to the frequency domain using FFT, obtaining the real part Re(X) and the imaginary part Im(X). By concatenating these two parts along the channel dimension, we obtain a more comprehensive frequency domain feature representation. Then, we learn from these features using a convolutional block, and finally transform the data back to the spatial domain using Inverse Fourier Transform (IFFT) to obtain feature X′. Furthermore, we apply residuals to the input feature X to obtain feature X′. f This further enhances the model's feature representation capabilities.

[0013] Re(X), Im(X) = FFT(X)

[0014] X′=IFFT(Conv(Cat[Re(X),Im(X)]))

[0015]

[0016] Where Gat[·,·] represents the concatenation operation along the channel dimension. This indicates element-wise addition. σ represents the Sigmoid function. Conv(·) represents a convolution block.

[0017] The obtained features are further processed through convolutional blocks to extract local information and provide a more complete feature representation F. Specifically, by capturing local features using convolution operations in the spatial domain and capturing global features through Fourier transform in the frequency domain, this fusion method can simultaneously capture both local details and global attributes of the image. This complementary method enhances the detection capability for small targets and improves the overall resolution capability for complex scenes.

[0018] 2. Feature enhancement resulting from network reconstruction

[0019] Infrared images are characterized by low contrast, noise, and blur, making it difficult to identify and detect small targets. Therefore, infrared image enhancement is necessary to improve target visibility and image clarity. We introduce a reconstruction network branch, enabling the model to handle image reconstruction and segmentation tasks simultaneously. By establishing a joint learning framework for reconstruction and segmentation, the model can exchange features and learn more useful representations for target detection.

[0020] The goal of image reconstruction is to obtain higher-quality, clearer, and more compact useful images by processing, repairing, and improving existing image data. Introducing a reconstruction network branch enhances infrared images, improving target visibility and image clarity, making small targets easier to identify and detect. Furthermore, the reconstruction network branch provides richer and more accurate image representations for the target detection branch. By learning and recovering the global structure and contextual information of the image, the reconstruction network helps the detection network better understand the location, shape, and features of targets in the image, thereby improving detection accuracy.

[0021] Specifically, such as Figure 2 As shown, we use a ResNet18-based UNet as the backbone of the reconstruction network, and employ CSAM (Spatial Channel Attention Module) between each convolutional block to integrate and enhance the acquired features. The raw infrared image, after preprocessing, serves as the input to the reconstruction network. The features L during the encoding process... i_0 It can be represented as:

[0022] L i_0 =ρ max (CSAM(Conv(L i-1_0 )))

[0023] Where, ρ max (·) denotes a max pooling operation with a stride of 2. CSAM represents the spatial channel attention module. Conv(·) represents a convolutional block in the backbone network.

[0024] During the decoding stage, deep features are upsampled and concatenated with the features from the encoding stage via ordinary skip connections, serving as input to the decoded block. The features from the decoding stage can be represented as:

[0025] L i_1 =CSAM(Conv(Cat[L i_0 ,μ(L i+1_1 )]))

[0026] Where Cat[·,·] represents the splicing operation along the channel dimension, and μ(·) represents the upsampling operation with a sampling ratio of 2.

[0027] For the detection network, the encoding part consists of regular convolutional blocks and FFT hybrid blocks. The output of the FFT hybrid block is represented as F... i Due to the output L from the corresponding stage of the reconstruction network i_0 This provides additional information to the detection network, thus increasing the final output feature D at each stage. i_0 It can be represented as:

[0028] D i_0 =ρ max (Cono(Cat[D i-1_0 L i_0 F i ]))

[0029] This method combines convolutional features, frequency domain properties, and reconstruction information, enhancing the encoder's ability to extract features from the input image.

[0030] 3. Cross-layer feature adaptive selection method

[0031] During the encoding phase, low-level features typically contain rich object details that are not readily apparent in high-level features. To address this issue, we propose a Cross-Layer Feature Adaptive Selection (CFAS) method, which refines high-level features by adaptively selecting the fusion weights of deep and shallow features, thereby facilitating shape and edge detection.

[0032] like Figure 3 As shown, for module input, low-level features First, a 1×1 convolution operation is used to adjust its channel dimensions so that they align with the high-level features. Matching. Then, the adjusted low-level features X... l With high-level characteristics X hElement-wise summation yields a fused feature, which serves as input to the channel attention (ca) module. The generated feature X... c Next, the channel attention weights f for high-level and low-level features are obtained using the sigmoid function. ca and 1-f ca These correspond to the characteristics of high-level and low-level areas, respectively.

[0033] f ca =Sigmoid(ca(X))

[0034] The output X of the channel attention module c The input is fed into the spatial attention (sa) module, and the result X is... s After sigmoid mapping, the weights f of high-level and low-level features in the spatial dimension are obtained. sa and 1-f sa :

[0035] f sa =Sigmoid(sa(X) c ))

[0036] Finally, the obtained weights are used to apply the weights to the low-level features X respectively. l and high-level characteristics X h The weighted values ​​are then processed, and residual operations are performed on the weighted results. Finally, the results are summed along the channel dimension to obtain the final enhanced feature X′:

[0037]

[0038] in, and These represent element-wise addition and element-wise multiplication, respectively.

[0039] Most existing feature fusion methods lack the flexibility to adjust the fusion weights of different features and typically only consider a single channel and spatial dimension. Compared to these methods, our approach offers several significant advantages. First, the weights of each feature are flexibly selected based on the characteristics of the feature itself. Second, it fully considers the spatial and channel information of the features. Finally, by incorporating a residual component, the robustness and generalization ability of the fusion are further enhanced.

[0040] 4. Loss Function

[0041] The loss function is a crucial component of model training, used to compare the difference between the real target and the predicted result. In Infrared Target Detection (IRSTD) tasks, IoU loss is commonly used. However, IoU loss focuses more on differences in the shape and area of ​​the predicted target, and is insensitive to the accuracy of the target's location. Therefore, we introduce the concept of localization calibration detection and design a loss function based on coordinate calibration. To address the insensitivity of the detection network to small target locations, we design a target-level evaluation metric for IRSTD tasks: coordinate calibration (CC) loss, and propose a cooperative two-stage training strategy aimed at calibrating and optimizing the predicted target's location. Specifically, as follows... Figure 4 As shown, we first obtain the centroid coordinates of the real and predicted targets. Since infrared images may contain multiple small infrared targets, we calculate the average of the centroid coordinates of these targets when multiple targets are present. Then, we calculate the positional information difference between the real and predicted targets and use it to further optimize the model.

[0042] First, the IoU loss, used in the IRSTD task, is a pixel-level evaluation metric that assesses the detector's ability to describe the target contour. IoU is calculated as the ratio of the area of ​​the intersection to the area of ​​the union between the predictions and the labels.

[0043]

[0044] in, A and A represent the predicted and actual target images, respectively. and These are the areas of the intersection and the union, respectively.

[0045] MSE loss is used for infrared image reconstruction tasks to measure the pixel-level difference between the model-reconstructed image and the original infrared image, specifically the average of the squared differences between the predicted and true values.

[0046]

[0047] in, and I i Here, ...

[0048]

[0049] Where λ is used to adjust CCloss The proportion of the total loss, where m is the number of targets (when there are multiple targets). These are the centroid coordinates of the actual target.

[0050] The specific implementation of the two-stage training strategy is shown below:

[0051]

[0052] Where n represents the current training round, and N is the round threshold for switching the loss function. Attached Figure Description

[0053] Figure 1 A schematic diagram of an infrared small target detection model based on a diversified feature capture and coordination network for the invention;

[0054] Figure 2 A schematic diagram of the branch structure for infrared image reconstruction in the model of the invention;

[0055] Figure 3 A schematic diagram of an adaptive selection method for cross-layer features;

[0056] Figure 4 This is a schematic diagram of coordinate calibration loss.

[0057] Figure 5 The visualization results of the method and comparison method of the present invention in various typical scenarios selected in the dataset;

[0058] Figure 6 The three-dimensional visualization results of the method and comparison method of the present invention in various typical scenarios selected in the dataset; Detailed Implementation

[0059] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0060] Step 1: Preprocess the data used;

[0061] Step 2: Input the preprocessed data into the target detection branch and the reconstruction branch for feature extraction;

[0062] Step 3: Input the features of the target detection branch into the cross-layer feature adaptive selection method module;

[0063] Step 4: Obtain the results of the reconstructed branches, and fuse the outputs at different scales to obtain the final prediction map;

[0064] Step 5: Calculate the loss using the loss function;

[0065] Furthermore, step 1 specifically includes the following steps:

[0066] Step 1-1: The images in the infrared small target dataset vary in size. To ensure consistency of model input, improve computational efficiency and memory utilization, simplify data augmentation, and facilitate model training, we uniformly cropped and scaled the infrared images to a size of 256×256, setting the values ​​for the augmented areas to 0.

[0067] Steps 1-2: Then, normalize the image and perform data augmentation operations, including random image flipping, blurring, etc.

[0068] Furthermore, step 2 specifically includes the following steps:

[0069] Step 2-1: As Figure 1 The diagram shows an infrared small target detection model based on the invention's diversified feature capture and coordination network. The implementation process of the infrared image reconstruction branch structure is described in detail below. Figure 2 As shown, before the infrared image is input for image reconstruction, the image is first processed by adding Gaussian noise of random size. The processed image is used as the input for the reconstruction branch, while the original image is used as the input for the target detection branch.

[0070] Step 2-2: We designed a hybrid encoding block based on FFT to effectively extract local and global features. The specific architecture of the encoding block is as follows: Figure 1 As shown in (a) and (b), we transform the input feature X to the frequency domain using FFT, obtaining the real part Re(X) and the imaginary part Im(X). By concatenating these two parts along the channel dimension, we obtain a more comprehensive frequency domain feature representation. Then, we learn from these features using a convolutional block, and finally transform the data back to the spatial domain using Inverse Fourier Transform (IFFT) to obtain feature X′. Furthermore, we apply residuals to the input feature X to obtain feature X′. f This further enhances the model's feature representation capabilities.

[0071] Re(X), Im(X) = FFT(X)

[0072] X′=IFFT(Cone(Cat[Re(X),Im(X)]))

[0073]

[0074] Cat[·,·] represents the concatenation operation along the channel dimension. This indicates element-wise addition. σ represents the Sigmoid function. Conv(·) represents a convolution block.

[0075] Steps 2-3: By learning and recovering the global structure and contextual information of the image, the reconstruction network can help the detection network better understand the location, shape and features of the target in the image, thereby improving the detection accuracy.

[0076] Specifically, such as Figure 2 As shown, we use a ResNet18-based UNet as the backbone of the reconstruction network, and employ CSAM (Spatial Channel Attention Module) between each convolutional block to integrate and enhance the acquired features. The raw infrared image, after preprocessing, serves as the input to the reconstruction network. The features L during the encoding process... i_0 It can be represented as:

[0077] L i_0 =ρ max (CSAM(Conv(L i-1_0 )))

[0078] Where, ρ max (·) denotes a max pooling operation with a stride of 2. CSAM represents the spatial channel attention module. Conv(·) represents a convolutional block in the backbone network.

[0079] During the decoding stage, deep features are upsampled and concatenated with the features from the encoding stage via ordinary skip connections, serving as input to the decoded block. The features from the decoding stage can be represented as:

[0080] L i_1 =CSAM(Conv(Cat[L i_0 ,μ(L i+1_1 )]))

[0081] Where Cat[·,·] represents the splicing operation along the channel dimension, and μ(·) represents the upsampling operation with a sampling ratio of 2.

[0082] For the detection network, the encoding part includes regular convolutional blocks and FFT hybrid blocks. The output of the FFT hybrid block is represented as F... i Due to the output L from the corresponding stage of the reconstruction network i_0 This provides additional information to the detection network, thus increasing the final output feature D at each stage. i_0 It can be represented as:

[0083] D i_0 =ρ max (Conv(Cat[D i-1_0 L i_0 F i ]))

[0084] Furthermore, step 3 specifically includes the following steps:

[0085] Step 3-1: As Figure 3 As shown, for module input, low-level features First, a 1×1 convolution operation is used to adjust its channel dimensions so that they align with the high-level features. Matching. Then, the adjusted low-level features X... l With high-level characteristics X h Element-wise summation yields a fused feature, which serves as input to the channel attention (ca) module. The generated feature X... c Next, the channel attention weights f for high-level and low-level features are obtained using the sigmoid function. ca and 1-f ca These correspond to the characteristics of high-level and low-level areas, respectively.

[0086] f ca =Sigmoid(ca(X))

[0087] Step 3-2: Output X of the channel attention module c The input is fed into the spatial attention (sa) module, and the result X is... s After sigmoid mapping, the weights f of high-level and low-level features in the spatial dimension are obtained. sa and 1-f sa :

[0088] f sa =Sigmoid(sa(X) c ))

[0089] Step 3-3: Finally, use the obtained weights to apply them to the low-level feature X. l and high-level characteristics X h The weighted values ​​are then processed, and residual operations are performed on the weighted results. Finally, the results are summed along the channel dimension to obtain the final enhanced feature X′:

[0090]

[0091] in, and These represent element-wise addition and element-wise multiplication, respectively.

[0092] Furthermore, step 4 specifically includes the following steps:

[0093] The features obtained in the above process are upsampled to the same size as the original image, and then a 1×1 convolutional layer is used to transform the channel size of the feature map to obtain a feature map with 3 channels. The global robust feature map G is calculated using the following formula:

[0094]

[0095] Finally, channel transformation is performed on the obtained robust feature map to generate the final prediction map.

[0096] Furthermore, step 5 specifically includes the following steps:

[0097] like Figure 4 As shown, we first obtain the centroid coordinates of the real and predicted targets. Since infrared images may contain multiple small infrared targets, we calculate the average of the centroid coordinates of these targets when multiple targets are present. Then, we calculate the positional information difference between the real and predicted targets and use it to further optimize the model. We use the intersection-union loss to predict the area difference between the real and predicted targets. The specific implementation of the two-stage training strategy is shown below:

[0098]

[0099] Where n represents the current training round, and N is the round threshold for switching the loss function.

[0100] We validate the state-of-the-art and effectiveness of our method using the publicly available NUAA, NUDT, and IRSTD-1k datasets from ACM Net, DNANet, and ISNet methods. The NUAA dataset contains 427 images from hundreds of real-world natural scenes, making it one of the most popular single-frame IRSTD datasets. The NUDT dataset contains 1327 images of more challenging scenes, such as multi-object, point-object, and dark-object scenes. The IRSTD-1k dataset contains 1000 infrared images from the real world, featuring targets of different types and sizes, complex scenes, and severe clutter and noise. The selected datasets ensure diversity and cover the various situations and challenges the model might face, enhancing the evaluation of the model's generalization ability.

[0101] In the model-driven approach, Top-hat and Max-median were selected respectively. In the data-driven approach, ACMNet, AGPCNet, DNANet, and DCFRNet were selected for comparison with the method of this invention. The resulting visualization results in typical scenarios and 3D visualization results are shown below. Figure 4 , Figure 5 As shown in Table 1, this invention uses mIoU (Intersection over Union) to quantitatively evaluate and analyze the method of this invention and several other infrared small target detection methods. The mIoU value ranges from 0 to 1; the closer the value is to 1, the better the performance of the detection method. The results are shown in Table 1. The experimental results show that the detection performance of the method of this invention is significantly better than the comparative algorithms, thus verifying the superiority of the method of this invention.

[0102] Table 1 Comparison of mIoU for Infrared Small Target Detection

[0103]

[0104] The results are shown in Table 1. The experimental results demonstrate that the detection performance of the method described in this invention is significantly better than the comparative algorithm, thus verifying the superiority of the method presented in this invention.

Claims

1. An infrared small target detection method based on diversified feature learning and coordination, characterized in that: Includes the following steps: S1. Construct a diverse feature capture and coordination network, which learns and coordinates diverse features through multi-path encoding; A global feature extraction branch composed of FFT hybrid coding blocks is constructed to capture macroscopic features of infrared images and provide extensive contextual information for target detection. S2. Construct an infrared image reconstruction branch that runs in parallel with the detection branch, preserving small target information and reducing feature loss through complementary context encoding; S3, Autonomous selection of cross-layer features: Cross-layer feature fusion is performed during the decoding stage to effectively integrate and coordinate deep semantic information and shallow detail features; S4. Introduce the concept of positioning calibration detection and propose a coordinate calibration loss function that focuses on optimizing the accuracy of the target position. At the same time, adopt a matching two-stage training strategy to accurately capture and correct the target position. Step S1 includes the following steps: S1.1 Designs an FFT-based hybrid coding block to effectively extract local and global features; the input feature X is transformed to the frequency domain using FFT to obtain the real part Re(X) and the imaginary part Im(X). By concatenating these two parts along the channel dimension, a more comprehensive frequency domain feature representation is obtained. Re(X), Im(X) = FFT(X) S1.2 uses a convolutional block to learn from these features, and finally transforms the data back to the spatial domain through inverse Fourier transform (IFFT) to obtain feature X′; X′=IFFT(Conv(Cat[Re(X),Im(X)])) S1.3 Apply residuals to the input feature X to obtain feature X. f ; Where Cat[·,·] represents the concatenation operation along the channel dimension. This indicates element-wise addition, σ represents the Sigmoid function, and Conv(·) represents a convolution block.

2. The infrared small target detection method based on diversified feature learning and coordination according to claim 1, characterized in that: Step S2 includes the following steps: S2.1 By learning and recovering the global structure and contextual information of an image, the reconstruction network can help the detection network better understand the location, shape, and features of targets in the image, thereby improving detection accuracy; UNet, based on ResNet18, is used as the backbone of the reconstruction network, and a CSAM spatial channel attention module is used between each convolutional block to integrate and enhance the acquired features. The original infrared image, after preprocessing, is used as the input to the reconstruction network, encoding the features L... i_0 Represented as: L i_0 =ρ max (CSAM(Conv(L i-1_0 ))) Where, ρ max (·) represents a max pooling operation with a step size of 2, and CSAM represents the spatial channel attention module; S2.2 In the decoding stage, the depth features are upsampled and connected together with the features from the encoding stage through ordinary skip connections, serving as the input to the decoding block. The features in the decoding stage are represented as follows: IT i_1 =CSAM(Conv(Cat[L i_0 ,μ(L i+1_1 )])) Where μ(·) represents the upsampling operation with a sampling ratio of 2; S2.3 For the detection network, the encoding part includes regular convolutional blocks and FFT hybrid blocks. The output of the FFT hybrid block is represented as F... i The final output feature D i_0 Represented as: D i_0 =ρ max (Conv(Cat[D i-1_0 ,L i_0 ,F i ]))。 3. The infrared small target detection method based on diversified feature learning and coordination according to claim 1, characterized in that: Step S3 includes the following steps: S3.1 For module input, low-level features First, a 1×1 convolution operation is used to adjust its channel dimensions so that they align with the high-level features. match; S3.2 will adjust the low-level feature X l With high-level characteristics X h Element-wise addition yields a fused feature, which serves as the input to the channel attention (ca) module, generating feature X. c Next, the channel attention weights f for high-level and low-level features are obtained using the sigmoid function. ca and 1-f ca These correspond to high-level and low-level features, respectively: f ca =Sigmoid(ca(X)) The output X of the S3.3 channel attention module c The input is fed into the spatial attention module (sa), and the result is X. s After sigmoid mapping, the weights f of high-level and low-level features in the spatial dimension are obtained. sa and 1-f sa : f sa =Sigmoid(sa(X c )) S3.4 Finally, the obtained weights are used to apply the weights to the low-level features X respectively. l and high-level characteristics X h The weighted values ​​are then processed, and residual operations are performed on the weighted results. Finally, the results are summed along the channel dimension to obtain the final enhanced feature X′: in, and These represent element-wise addition and element-wise multiplication, respectively.

4. The infrared small target detection method based on diversified feature learning and coordination according to claim 1, characterized in that: The specific steps of step S4 are as follows: First, obtain the centroid coordinates of the real and predicted targets, and calculate the average of these centroid coordinates. Then, calculate the positional information difference between the real and predicted targets, and use it to further optimize the model. Use the intersection-union loss to predict the area difference between the real and predicted targets. The specific implementation of the two-stage training strategy is shown below: Where n represents the current training round, and N is the round threshold for switching the loss function; This represents the intersection and union ratio loss, where Let A and B represent the predicted target map and the actual target map, respectively. and These are the areas of the intersection and the union, respectively. This represents the average of the squared differences between the predicted and actual values; where, and I i These are the values ​​of the i-th pixel in the reconstructed image and the original image, respectively, and M is the total number of pixels in the image; This represents the coordinate calibration loss; where λ is used to adjust the CC. loss The proportion of the total loss, where m is the number of targets. These are the centroid coordinates of the actual target; The loss function of the two-stage training strategy is used to calculate the detection branch loss and the image reconstruction branch loss for model training.

Images