An infrared small target detection method based on U-shaped network with pulse coupled filter fusion
By using a U-shaped network that integrates pulse coupling filtering and cross-layer feature enhancement modules, the problems of background interference and semantic alignment in infrared small target detection are solved, achieving a balance between high detection accuracy and model complexity, and making it suitable for infrared small target detection in complex scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2026-04-22
- Publication Date
- 2026-06-19
AI Technical Summary
Existing infrared small target detection methods struggle to effectively filter background and clutter information in complex environments, and deep feature semantic information is difficult to guide shallow details, making it difficult to balance detection accuracy and model complexity.
A U-shaped network with fused pulse-coupled filtering is adopted. By designing an infrared small target pulse-coupled filtering module to filter clutter information, and using a cross-layer feature enhancement module to achieve semantic alignment between deep semantics and shallow details, the detection accuracy and robustness are improved in synergy.
While maintaining a lightweight design, it significantly improves the accuracy and robustness of infrared small target detection, achieving a good balance between detection accuracy and model complexity, and is suitable for resource-constrained scenarios.
Smart Images
Figure CN122243799A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a U-shaped network infrared small target detection method that integrates pulse coupling filtering. Background Technology
[0002] Infrared small target detection aims to accurately identify tiny, weak-signal potential targets from complex infrared backgrounds. It is a key technology in military and civilian fields such as precision guidance, military reconnaissance, and maritime search. Compared with visible light images, infrared images are more significantly affected by atmospheric attenuation, thermal noise, and interference from complex terrain features. Targets often lack clear outlines and texture features, appearing only as faint bright spots of a few to tens of pixels. In addition, the highly dynamic background and dense clutter in real-world scenes make targets easily submerged in noise, causing traditional detection methods to degrade sharply under low signal-to-noise ratio conditions.
[0003] In recent years, deep learning methods based on U-Net have made some progress, but they have overlooked the fact that the features extracted by the encoder are often mixed with a lot of background and clutter information. This information is amplified in skip connections and decoders, which seriously interferes with detection performance. The deep features of the network are rich in semantics but lack spatial details, while the shallow features are complete in detail but lack semantic information. Current mainstream networks often find it difficult to achieve an effective balance between detection accuracy and model complexity. Summary of the Invention
[0004] This invention addresses the shortcomings of existing technologies by proposing a U-shaped network infrared small target detection method that integrates pulse-coupled filtering. This method aims to effectively filter clutter and background information from the decoder to the encoder, effectively utilize the semantic information of deep features to supplement the semantic information of shallow features, and achieve accurate detection of infrared small targets with minimal complexity. To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a U-shaped network infrared small target detection method incorporating pulse coupling filtering, characterized by the following steps: Step 1: Obtain the preprocessed infrared image dataset ;in, Indicates the first k Zhang preprocessed infrared image W and H These represent the width and height of each infrared image, respectively. The number of channels in each infrared image, and =1; K Represents the total number of infrared images; let The real infrared target image is ; right Perform data augmentation to obtain the first k Zhang enhanced infrared image Thus, the enhanced infrared image dataset is obtained. ; Step 2: Construct an infrared small target detection network, including: a residual structure-based encoding module, a bottleneck layer module, an infrared small target filtering module, and a prediction module, and perform... Processing yields the first... i The first layer k Zhang Hongguang Significance Prediction Chart and the k Infrared splicing saliency prediction map ; Step 3: Construct the total loss of the infrared small target detection network using equation (11). : (11) In equation (19), Indicates the first k In the infrared saliency prediction map sequence, the first... The loss weights of each significant prediction plot, Indicates the first The loss weights of each infrared stitching saliency prediction map, Represents the binary cross-entropy loss function; Step 4: Iteratively train the infrared small target detection network using the Adam optimizer until the total loss is reached. The process continues until convergence, thus obtaining a trained infrared small target detection model for detecting infrared small targets.
[0005] The U-shaped network infrared small target detection method based on fused pulse coupling filtering described in this invention is characterized by including the following steps: Step 2.1, the encoding module includes M Layered structure, wherein, the first i The layer structure is obtained using equation (1). The i The first layer k Infrared coded features ,in, express The number of channels, It is a constant; (1) In equation (1), Indicates that the convolution kernel is × Two-dimensional convolution operation, Indicates that the convolution kernel is × Two-dimensional convolution operation, For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation; Indicates residual connection, express The i- The first floor k One infrared coded feature; , Different sizes of convolution kernels; Step 2.2, the bottleneck layer module uses equation (2) to process the first... M The first layer k Infrared coded features Processing yields the first... k Bottleneck layer decoding features ; (2) Step 2.3: The infrared small target filtering module consists of an aggregation module, a pulse coupling filtering module, a cross-layer feature enhancement module, and a decoder module, and performs filtering on the first... k Infrared coded feature sequence Features of the bottleneck layer Processing yields the first... k Decoded feature sequences ,in, Indicates the first i The first layer k One decoding feature; Step 2.4: Obtain the first result using equation (10). k Infrared splicing saliency prediction map : (10) In equation (18), Indicates that the convolution kernel is Two-dimensional convolution operation, This indicates splicing along the channel dimension. for Activation function.
[0006] Furthermore, step 2.3 includes the following steps: Step 2.3.1: The aggregation module uses formula (3) to... Processing yields the first... The first layer k One aggregation feature ,in, ; (3) In equation (3), This represents the convolutional attention mechanism. Indicates splicing along the channel dimension; when season ; Step 2.3.2: The pulse coupling filter module uses convolution kernels for... × Two-dimensional convolution operations Projecting, we obtain the first... The first layer k Projection features ; Initialize the current iteration count t =1 ， Initialize the first t- The first iteration i The first layer k Each filter feature = 0, no. t- The first iteration The first layer k Internal activity characteristics = 0, no. t- The first iteration The first layer k Threshold features = 1; thus, using equation (5) to... and Conduct the first t In the nth iteration, we obtain the nth t The iteration of the ... The first layer k Internal activity characteristics , No. t The iteration of the ... The first layer k Each filter feature and the t The iteration of the ... The first layer k Threshold features Thus complete N After the nth iteration, we obtain the nth... The iteration of the ... The first layer k Each filter feature : (5) In equation (5), Indicates the first t The iteration of the ... Internal activity characteristics of the layer Indicates the first t The iteration of the ... The filtering characteristics of the layer Indicates the first t The iteration of the ... Threshold features of the layer , as well as All are parameters to be learned. p For steep slope parameters, This represents the Sigmoid activation function. For max pooling operation, For average pooling operation, Indicates that the convolution kernel is × Depth convolution; Using equation (6), we obtain the first... The first layer k pulse coupling characteristics : (6) Step 2.3.3: The cross-layer feature enhancement module uses equation (7) to... and Processing yields the first... The first layer k One enhanced feature ,in, ; (7) In equation (7), Indicates standard channel attention. Indicates the first The first layer k An attention enhancement feature, Indicates the first The first layer k A localized refinement feature; This represents the bilinear interpolation operation. Indicates that the convolution kernel is × Depth convolution, and There are two parameters to be learned; when season ; Step 2.3.4: The decoder module uses equation (8) to... Processing yields the first... The first layer k Decoding features : (8) In equation (8), They represent the first The first layer k The first branch decoding feature and the first k The second branch decoding feature, Indicates that the convolution kernel is × convolution, Indicates efficient channel attention; Step 2.3.5: Obtain the first result through equation (9). i The first layer k A sequence of infrared saliency prediction maps : (9) In equation (9), Indicates that the convolution kernel is Two-dimensional convolution operation, This is a bilinear interpolation upsampling operation. for Activation function.
[0007] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.
[0008] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.
[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Traditional methods often rely on handcrafted features or single convolutional neural networks, making it difficult to adapt to complex scene changes and achieve a good balance between detection accuracy and robustness. They are prone to false negatives and missed detections in complex scenarios. With the rapid development of deep learning technology in recent years, significant progress has been made in infrared small target detection based on convolutional neural networks. Regarding improvements to the U-shaped architecture, researchers have mainly focused on skip connections and multi-scale feature fusion. However, they have overlooked the fact that the features extracted by the encoder often contain a large amount of background and clutter information, which is easily amplified in skip connections, severely interfering with detection performance. This invention focuses on the field of infrared small target detection and innovatively proposes a U-shaped network infrared small target detection method that integrates pulse-coupled filtering. By designing an infrared small target pulse-coupled filtering module, clutter and background information in the encoder and deep decoder features are filtered out, significantly enhancing the detection capability of the decoder and model for small targets.
[0010] 2. Recent deep learning-based methods often employ feature fusion to address the interaction between deep and shallow network layers. However, these methods mostly achieve cross-layer feature fusion through simple concatenation or addition, ignoring the inherent semantic gap between shallow and deep features. This results in the target response in the fused features being easily drowned out by background noise, and deep semantic information failing to effectively guide shallow details to focus on the target region. To address this issue, this invention further designs a cross-layer feature enhancement module. This module uses the semantic information of deep features to adaptively recalibrate shallow features. Through channel statistical fusion and local refinement mechanisms, it achieves semantic alignment and efficient fusion between deep semantics and shallow details, effectively compensating for the shortcomings of existing methods in providing insufficient semantic guidance for the target in cross-layer interactions. This further improves the model's detection accuracy and robustness for small, weak targets. The two modules work together to enhance the network's ability to perceive small infrared targets from the dimensions of background suppression and semantic alignment, respectively, achieving a good balance between detection accuracy and model complexity.
[0011] 3. Deep learning-based methods often significantly increase the number of parameters and computational load when improving detection accuracy, making it difficult to achieve an effective balance between detection performance and model complexity, thus limiting their deployment in resource-constrained scenarios. This invention, while maintaining a lightweight design, effectively controls the number of model parameters and floating-point operations through the synergistic effect of the pulse coupling filter module and the cross-layer feature enhancement module, significantly improving detection accuracy while achieving a good balance between detection accuracy and model complexity. It possesses strong engineering application value and promotion potential. Attached Figure Description
[0012] Figure 1 This is a flowchart of the U-shaped network infrared small target detection method integrating pulse coupling filtering in this invention; Figure 2 This is a flowchart of the pulse coupling filter module of the present invention; Figure 3 This is a flowchart of the cross-layer feature enhancement module in this invention; Figure 4 This is a comparison of the real infrared target image, the predicted image of the U-shaped network infrared small target detection algorithm with pulse coupling filtering fusion, and the real infrared target image in the verification experiment of this invention. Detailed Implementation
[0013] In this embodiment, an infrared small target detection method based on fuzzy logic reasoning is described, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the preprocessed infrared image dataset ;in, Indicates the first k Zhang preprocessed infrared imageW and H These represent the width and height of each infrared image, respectively. The number of channels in each infrared image, and =1; K Represents the total number of infrared images; let The real infrared target image is In this embodiment, the width of the preprocessed infrared image and the enhanced infrared image are... W =265, H =256.
[0014] right Perform data augmentation to obtain the first k Zhang enhanced infrared image Thus, the enhanced infrared image dataset is obtained. .
[0015] Step 2: Construct an infrared small target detection network, including: a residual structure-based encoding module, a bottleneck layer module, an infrared small target filtering module, and a prediction module, and perform... Processing yields the first... i The first layer k Zhang Hongguang Significance Prediction Chart and the k Infrared splicing saliency prediction map .
[0016] Step 2.1, the encoding module includes M Layered structure, wherein, the first i The layer structure is obtained using equation (1). The i The first layer k Infrared coded features ,in, express The number of channels, It is a constant; in this embodiment, M =4, r =32.
[0017] (1) In equation (1), This represents a 2D convolution with a 3×3 kernel and a stride of 1. This represents a 2D convolution with a 1×1 kernel and a stride of 1. For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation with a step size of 2; Indicates residual connection, express The i- The first floor k One infrared coded feature.
[0018] Step 2.2, the bottleneck layer module uses equation (2) to process the first... M The first layer k Infrared coded features Processing yields the first... k Bottleneck layer decoding features
[0019] (2) Step 2.3: The infrared small target filtering module consists of an aggregation module, a pulse coupling filtering module, a cross-layer feature enhancement module, and a decoder module, and performs filtering on the first... k Infrared coded feature sequence Features of the bottleneck layer Processing yields the first... k Decoded feature sequences ,in, Indicates the first i The first layer k One decoding feature.
[0020] Step 2.3.1: The aggregation module uses formula (3) to... Processing yields the first... The first layer k One aggregation feature ,in During initialization, when season ; (3) In equation (3), This represents the convolutional attention mechanism. This indicates splicing along the channel dimension.
[0021] Step 2.3.2, Pulse Coupling Filter Module, such as Figure 2 As shown, using convolution kernels as × Two-dimensional convolution operations Projecting, we obtain the first... The first layer k Projection features .
[0022] Initialize the current iteration count t =1 ， Initialize the first t- The first iteration M The first layer k Each filter feature = 0, no. t- The first iteration The first layer k Internal activity characteristics = 0, no. t- The first iteration The first layer k Threshold features = 1; thus, using equation (5) to... and Conduct the first t In the nth iteration, we obtain the nth t The iteration of the ... The first layer k Internal activity characteristics , No. t The iteration of the ... The first layer k Each filter feature and the t The iteration of the ... The first layer k Threshold features Thus complete N After the nth iteration, we obtain the nth... The iteration of the ... The first layer k Each filter feature In this example, : (5) In equation (5), Indicates the first t The iteration of the ... Internal activity characteristics of the layer Indicates the first t The iteration of the ... The filtering characteristics of the layer Indicates the first t The iteration of the ... Threshold features of the layer , as well as All are parameters to be learned. p For steep slope parameters, in this example p =10, This represents the Sigmoid activation function. This is a max pooling operation with a step size of 2. For average pooling operation, This represents a depthwise convolution with a kernel size of 3×3 and a stride of 1.
[0023] Using equation (6), we obtain the first... The first layer k pulse coupling characteristics : (6) Step 2.3.3, Cross-layer feature enhancement module, such as Figure 3 As shown, by using equation (7) and Processing yields the first... The first layer k One enhanced feature ,in During initialization, when season ; (7) In equation (7), Indicates standard channel attention. Indicates the first The first layer k An attention enhancement feature, Indicates the first The first layer k A localized refinement feature; This represents the bilinear interpolation operation. This represents a depthwise convolution with a 1×1 kernel and a stride of 1. and There are two parameters to be learned.
[0024] Step 2.3.4: The decoder module uses equation (8) to... Processing yields the first... The first layer k Decoding features : (8) In equation (8), They represent the first The first layer k The first branch decoding feature and the first k The second branch decoding feature, This represents a 2D convolution with a 5×5 kernel and a stride of 1. This indicates efficient channel attention.
[0025] Step 2.3.5: Obtain the first result through equation (9). i The first layer k A sequence of infrared saliency prediction maps : (9) In equation (9), Indicates that the convolution kernel is Two-dimensional convolution operation, This is a bilinear interpolation upsampling operation. for Activation function.
[0026] Step 2.4: Obtain the first result using equation (10). k Infrared splicing saliency prediction map : (10) Step 3: Construct the total loss of the infrared small target detection network using equation (11). : (11) In equation (11), Indicates the first k In the infrared saliency prediction map sequence, the first... The loss weights of each significant prediction plot, Indicates the first The loss weights of each infrared stitching saliency prediction map, This represents the binary cross-entropy loss function.
[0027] Step 4: Iteratively train the infrared small target detection network using the Adam optimizer until the total loss is reached. The process continues until convergence, thus obtaining a trained infrared small target detection model for detecting infrared small targets.
[0028] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.
[0029] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.
[0030] To verify the effectiveness of this invention, this embodiment applies the proposed method to an infrared small target detection task, and selects four challenging infrared small target scene images for visualization analysis. The results are as follows: Figure 4 As shown. In Figure 4In the diagram, each row contains: the original infrared image, the predicted image generated by this method, and the original target annotation. To clearly display the details of small targets, the target area in each image is magnified and placed in the lower right corner. The target position in the original image is marked with a red circle, and the magnified area is indicated by a red border for easy observation. As can be seen from the diagram, the target predicted by this method closely matches the actual annotation in both position and shape. This visualization result intuitively demonstrates the accuracy and reliability of the method for detecting small infrared targets in complex scenes.
[0031] To objectively evaluate the performance of the method of this invention, this embodiment compares it with current mainstream infrared small target detection algorithms on the same test set, including ALCNet, RDIAN, MSHNet, SDSNet, L²SKNet, and HAFNet. Five commonly used evaluation metrics in this field were selected: Intersection over Union (IoU), Normalized Intersection over Union (nIoU), and Detection Rate. False alarm rate And F1 score. Among them, IoU, nIoU, A higher F1 score indicates better overall model performance; The smaller the value, the lower the false positive rate and the better the performance of the model.
[0032] Table 1. Performance comparison with other algorithms in infrared small target detection accuracy.
[0033] As can be seen from the results in Table 1, the method of the present invention achieves good results in IoU, nIoU, Both F1 and F1 achieved the highest values, while maintaining the minimum. The above results fully demonstrate that the method of the present invention has superior overall performance and higher detection reliability in infrared small target detection tasks.
Claims
1. A U-shaped network infrared small target detection method incorporating pulse coupling filtering, characterized in that, Includes the following steps: Step 1: Obtain the preprocessed infrared image dataset ;in, Indicates the first k Zhang preprocessed infrared image W and H These represent the width and height of each infrared image, respectively. The number of channels in each infrared image, and =1; K Represents the total number of infrared images; let The real infrared target image is ; right Perform data augmentation to obtain the first k Zhang enhanced infrared image Thus, the enhanced infrared image dataset is obtained. ; Step 2: Construct an infrared small target detection network, including: a residual structure-based encoding module, a bottleneck layer module, an infrared small target filtering module, and a prediction module, and perform... Processing is performed to obtain the first... i The first layer k Zhang Hongguang Significance Prediction Chart and the k Infrared splicing saliency prediction map ; Step 3: Construct the total loss of the infrared small target detection network using equation (11). : (11) In equation (11), Indicates the first k In the infrared saliency prediction map sequence, the first... The loss weights of each significant prediction plot, Indicates the first The loss weights of each infrared stitching saliency prediction map, Represents the binary cross-entropy loss function; Step 4: Iteratively train the infrared small target detection network using the Adam optimizer until the total loss is reached. The process continues until convergence, thus obtaining a trained infrared small target detection model for detecting infrared small targets.
2. The method for detecting small infrared targets using a U-shaped network with fused pulse coupling filtering according to claim 1, characterized in that, Includes the following steps: Step 2.1, the encoding module includes M Layered structure, wherein, the first i The layer structure is obtained using equation (1). The i The first layer k Infrared coded features ,in, express The number of channels, It is a constant; (1) In equation (1), Indicates that the convolution kernel is × Two-dimensional convolution operation, Indicates that the convolution kernel is × Two-dimensional convolution operation, For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation; Indicates residual connection, express The i- The first floor k One infrared coded feature; , Different sizes of convolution kernels; Step 2.2, the bottleneck layer module uses equation (2) to process the first... M The first layer k Infrared coded features Processing is performed to obtain the first... k Bottleneck layer decoding features ; (2) Step 2.3: The infrared small target filtering module consists of an aggregation module, a pulse coupling filtering module, a cross-layer feature enhancement module, and a decoder module, and performs filtering on the first... k Infrared coded feature sequence Features of the bottleneck layer Processing is performed to obtain the first... k Decoded feature sequences ,in, Indicates the first i The first layer k One decoding feature; Step 2.4: Obtain the first result using equation (10). k Infrared splicing saliency prediction map : (10) In equation (18), Indicates that the convolution kernel is Two-dimensional convolution operation, This indicates splicing along the channel dimension. for Activation function.
3. The U-shaped network infrared small target detection method with fused pulse coupling filtering according to claim 2, characterized in that, Step 2.3 includes the following steps: Step 2.3.1: The aggregation module uses formula (3) to... Processing is performed to obtain the first... The first layer k One aggregation feature ,in, ; (3) In equation (3), This represents the convolutional attention mechanism. Indicates splicing along the channel dimension; when season ; Step 2.3.2: The pulse coupling filter module uses convolution kernels for... × Two-dimensional convolution operations Projecting, we obtain the first... The first layer k Projection features ; Initialize the current iteration count t =1 ， Initialize the first t- The first iteration i The first layer k Each filter feature = 0, no. t- The first iteration The first layer k Internal activity characteristics = 0, no. t- The first iteration The first layer k Threshold features = 1; thus, using equation (5) to... and Conduct the first t In the nth iteration, we obtain the nth t The iteration of the ... The first layer k Internal activity characteristics , No. t The iteration of the ... The first layer k Each filter feature and the t The iteration of the ... The first layer k Threshold features Thus complete N After the nth iteration, we obtain the nth... The iteration of the ... The first layer k Each filter feature : (5) In equation (5), Indicates the first t The iteration of the ... Internal activity characteristics of the layer Indicates the first t The iteration of the ... The filtering characteristics of the layer Indicates the first t The iteration of the ... Threshold features of the layer , as well as All are parameters to be learned. p For steep slope parameters, This represents the Sigmoid activation function. For max pooling operation, For average pooling operation, Indicates that the convolution kernel is × Depth convolution; Using equation (6), we obtain the first... The first layer k pulse coupling characteristics : (6) Step 2.3.3: The cross-layer feature enhancement module uses equation (7) to... and Processing is performed to obtain the first... The first layer k One enhanced feature ,in, ; (7) In equation (7), Indicates standard channel attention. Indicates the first The first layer k An attention enhancement feature, Indicates the first The first layer k A localized refinement feature; This represents the bilinear interpolation operation. Indicates that the convolution kernel is × Depth convolution, and There are two parameters to be learned; when season ; Step 2.3.4: The decoder module uses equation (8) to... Processing is performed to obtain the first... The first layer k Decoding features : (8) In equation (8), They represent the first The first layer k The first branch decoding feature and the first k The second branch decoding feature, Indicates that the convolution kernel is × convolution, Indicates efficient channel attention; Step 2.3.5: Obtain the first result through equation (9). i The first layer k A sequence of infrared saliency prediction maps : (9) In equation (9), Indicates that the convolution kernel is Two-dimensional convolution operation, This is a bilinear interpolation upsampling operation. for Activation function.
4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.
5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-3.