Infrared small target detection method based on visual image and frequency analysis coupling

Abstract

The invention discloses an infrared small target detection method based on coupling of a visual image and frequency analysis. The infrared small target detection method comprises the following steps: step 1, carrying out normalization and random overturning and rotation enhancement on an input infrared image; 2, extracting multi-scale features of the enhanced infrared image by using a residual block and a pooling layer; 3, splicing the multi-scale feature maps extracted by the encoder, and inputting the spliced multi-scale feature maps into a space-frequency coupling topology module to construct a long-range spatial dependency relationship and enhance frequency components related to a target; step 4, encoding the output enhanced features and the features of an upper-level decoder in a frequency domain, generating a fusion weight by using a channel attention mechanism, and fusing the features by using a mode of adding and averaging; 5, calculating loss and optimizing network parameters for decoding output of each level by adopting a deep supervision strategy; and 6, outputting a detection result. According to the method, the infrared image can be effectively and globally modeled, background clutter interference is reduced, and accurate detection of the target is realized.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to an infrared small target detection method that integrates graph structure learning and frequency domain analysis. Background Technology

[0002] Infrared small target detection has significant applications in military reconnaissance, maritime rescue, and security monitoring. However, infrared small targets typically have characteristics such as small pixel ratio, low signal-to-clutter ratio, and lack of significant structural features, making them susceptible to interference from complex background clutter and sensor noise. Traditional methods rely on manually designed features and parameters, resulting in poor generalization ability.

[0003] In recent years, although deep learning methods based on U-Net have made some progress, they still have problems such as insufficient global modeling ability, difficulty in capturing long-distance dependencies, simple feature fusion methods that fail to effectively combine high and low layer semantics, and lack of effective use of frequency domain information, making it difficult to suppress background clutter. Summary of the Invention

[0004] The present invention addresses the shortcomings of the prior art by proposing an infrared small target detection method based on the coupling of visual image and frequency analysis. This method can effectively model infrared images globally and reduce background clutter interference, thereby achieving accurate target detection. To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The infrared small target detection method based on the coupling of visual image and frequency analysis of the present invention is characterized by the following steps: Step 1: Obtain the preprocessed infrared image dataset ;in, Indicates the first k Zhang is a preprocessed infrared image, and K represents the total number of infrared images; let The real infrared target image is ;in, W and H These represent the width and height of each infrared image, respectively, and 1 represents the number of channels in each infrared image; 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: an encoder module based on residual blocks, a space-frequency coupled topology module, a decoding module, and a prediction module, and perform... Processing is performed to obtain the first... k Zhang Hongguang Significance Prediction Chart ; Step 2.1, the Encoder module includes... M The structure consists of layers, each layer comprising a residual block and a max-pooling layer, and these layers are sequentially processed. After processing, the processed infrared coded feature sequence is normalized in the spatial dimension, and then the normalized infrared coded feature sequence is concatenated in the channel dimension to form the first... k Infrared global features ; Step 2.2: The space-frequency coupled topology module consists of a visual interactive graph module and a dual-branch spectrum-enhanced frequency control feedforward network, and... Processing is performed to obtain the first... k Infrared Enhancement Feature Sequence ; Step 2.3, Decoding module Processing is performed to obtain the first... k Infrared decoding feature sequence ;in, Indicates the first i One infrared decoding feature; Step 2.4, the prediction module... Processing is performed to obtain the first... k Infrared splicing saliency prediction map ; Step 3: Construct the total loss of the infrared small target detection network using equation (1). ; (1) In equation (1), Indicates 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; Indicates the first k One infrared saliency prediction map; 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 infrared small target detection method based on the coupling of visual image and frequency analysis described in this invention is also characterized in that, in section 2.1, it includes: Step 2.1.1 Passing through in sequence M After processing the layer structure, the first layer is obtained. kInfrared coded feature sequence ,in, Indicates the first i Infrared encoded features output by the layer structure Indicates the first i The number of channels for layer infrared coded features; Represents a constant; Step 2.1.2: Use bilinear interpolation to perform spatial interpolation. After normalization, we obtain the first... k A normalized infrared coded feature sequence ,in, express The Middle i A normalized infrared coded feature; Step 2.1.3, on the channel dimension After concatenating each infrared coded feature in the image, we obtain the first... k Infrared global features ,in, The number of channels for the global infrared feature.

[0006] Furthermore, the first part of the Encoder module in section 2.1.1... i The layered structure is obtained using equation (2). : (2).

[0007] Furthermore, step 2 includes: Step 2.2.1, Visual Interaction Graph Module Processing is performed to obtain the first... k Infrared interactive graph features ; Step 2.2.2: The frequency-controlled feedforward network uses Fourier transform and gated spectral attention pairs... Processing is performed to obtain the first... k Infrared feedforward features ; Step 2.2.3, along the channel dimension Cut into M block, get the first k A set of blocks ,in, Indicates the first i Each block; Step 2.2.4: Use bilinear interpolation to... Length and width with Transform it to be consistent, and obtain the first... k A transformed block set ,in, express The transformed first i Each block; Step 2.2.5: Transform the block set With infrared coded feature sequences Add them according to their indices to get the first... k Infrared enhancement features ,in, ,express The Middle i One infrared enhancement feature.

[0008] Furthermore, step 2.2.1 includes: Step 2.2.1.1, will Evenly divided into P A set of blocks of the same size ,in, Indicates the first i One block; Step 2.2.1.2, Calculation and The Euclidean distance between other blocks in the middle is selected, and then... The closest Each block is connected to form a structure and The edges between the blocks, thus obtaining The edge connection relationships between all blocks in the array are used to generate the first... A binary adjacency matrix If there is an edge connecting the two blocks, then... The corresponding row and column element is set to "1"; otherwise, it is set to "0", thus determining the order based on the given information. and Construct the first k Individual diagram structure ; Step 2.2.1.3: Use the maximum relative graph convolution operator to pair Perform feature encoding to obtain the first k Interaction graph features .

[0009] Furthermore, step 2.2.2 includes: Step 2.2.2.1, along the channel dimension All are divided into the first infrared interactive image sub-features Second infrared interaction map features and respectively using a core size of and Grouped convolution pairs and Spatial local refinement is performed, resulting in the first infrared refined interactive map sub-feature. Second infrared refined interactive image features ; Step 2.2.2.2: Using equations (3) and (4) to... and The first spectral modulation coefficient is obtained through processing. Second spectrum modulation coefficient : (3) (4) In equations (3) and (4), rfft2 represents the Fast Fourier Transform; This indicates that the amplitude is taken, and MLP(·) represents a multilayer perceptron; This indicates channel attention gating. This represents the spectrum gating to be learned; Step 2.2.2.3: Obtain the first decoupling reconstruction feature using equations (5) and (6). Second decoupling reconstruction features ; (5) (6) In equations (5) and (6), This represents spatial attention gating, and rfft2 represents the Fast Fourier Transform. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. e It is a natural constant; Step 2.2.2.4: Use equations (7) and (8) to obtain the first... k feedforward features : (7) (8) In equations (7) and (8), Indicates the first k One aggregation feature, This indicates a splicing operation along the channel dimension. is the ReLU activation function; ECA represents the channel attention mechanism.

[0010] Furthermore, step 2.3 includes: when When =1, the first equation is obtained using equation (9). i- 1 infrared decoding feature : (9) In equation (9), Indicates that the convolution kernel is n × n Two-dimensional convolution operation, For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation; when When =1,2,…,M, the first... i Infrared enhancement features Will with the i- 1 infrared decoding feature By merging, we obtain the first... i A fused infrared feature And using equation (10) to obtain the first i Infrared decoding features Thus, the first k Infrared decoding feature sequence ; (10).

[0011] Furthermore, step 2.3 is a description of the first... Infrared decoding features Perform bilinear interpolation to obtain the first... Infrared interpolation features Therefore, using equations (11), (12), (13), and (14), we obtain : (11) (12) (13) (14) In equations (11)-(14), express n × n Two-dimensional convolution operation, ECA This represents channel attention, and rfft2 represents the Fast Fourier Transform. GAP This indicates the average pooling operation. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. It is a natural constant. This indicates taking its amplitude, and irfft2 represents the Fast Fourier Transform. Indicates by The generated weight map, Indicates by The generated weight map, express Activation function Indicates the first k A final weighted graph, express Activation function.

[0012] Furthermore, step 4 includes: Step 2.4.1: Use equation (15) to obtain the first... i Infrared saliency prediction map Thus, the first k Infrared saliency prediction map ; (15) In equation (17), This is a bilinear interpolation operation. Use the Sigmoid activation function; Step 2.4.2: Using equation (16) to obtain the first... k Infrared splicing saliency prediction map : (16) In equation (18), This indicates splicing along the channel dimension.

[0013] 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.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention focuses on the field of infrared small target detection, and innovatively proposes a detection method based on the coupling of visual graphs and frequency analysis. This effectively overcomes the performance bottlenecks of traditional infrared small target detection methods under extreme conditions such as low signal-to-noise ratio, complex backgrounds, and extremely small target sizes. Traditional methods often rely on handcrafted features or single convolutional neural networks, making it difficult to fully extract the spatial structural information and frequency domain features of weak targets in infrared images. This leads to frequent missed detections and false detections in scenarios with complex backgrounds and low target contrast. This invention, by constructing a space-frequency coupled topology module, deeply integrates visual interaction graph modeling with a spectrum enhancement mechanism. It fully utilizes the local structural relationships of the target in the spatial domain and the saliency response in the frequency domain, significantly improving the detectability and robustness of small targets, and providing more reliable technical support for key applications such as infrared early warning, nighttime surveillance, and military reconnaissance.

[0015] 2. This invention constructs an end-to-end infrared small target detection network, encompassing a residual coding module, a spatial-frequency coupled topology module, a decoding enhancement module, and a multi-scale prediction fusion mechanism, achieving high-precision, multi-scale, and robust detection of infrared small targets. Compared to traditional methods that rely solely on spatial convolution or single-scale features, this invention captures the long-range dependency between the target and the background through graph structure modeling, and combines Fourier transform and gated spectral attention mechanisms to effectively suppress background noise and enhance target response in the frequency domain. Simultaneously, the network introduces cross-scale feature fusion and attention weighting mechanisms in the decoding stage, further strengthening the model's ability to perceive small targets of different sizes. This structure not only achieves breakthroughs in detection accuracy but also demonstrates significant advantages in adaptability to complex environments, computational efficiency, and model generalization ability, laying a solid foundation for the engineering deployment and industrial application of infrared small target detection tasks.

[0016] 3. This invention features targeted optimizations in training strategies and loss design. Utilizing a deep supervision mechanism, it constructs joint loss constraints by combining layer-by-layer supervision of the saliency prediction map with overall supervision of the stitched saliency prediction map, achieving refined control over pixel-level localization accuracy for small infrared targets. Traditional methods often rely solely on single output supervision, making it difficult to fully exploit the network's discriminative capabilities at different abstraction levels, easily leading to blurred target edges or weak responses. This invention, by combining layer-by-layer supervision of the saliency prediction map with overall supervision of the stitched saliency prediction map, effectively enhances the model's consistent perception of target shape, location, and saliency, significantly reducing false alarm and false negative rates. Furthermore, the network training employs the Adam optimizer combined with a convergent loss design, greatly improving convergence speed while ensuring training stability. This enables the model to achieve rapid training and high-precision detection even on small datasets, providing an efficient and low-cost solution for scenarios where data acquisition is difficult and annotation costs are high in practical applications. Attached Figure Description

[0017] Figure 1 This is a flowchart of the infrared small target algorithm based on the coupling of visual graph and frequency analysis in this invention.

[0018] Figure 2 This is a flowchart of the space-frequency coupled topology module in this invention.

[0019] Figure 3 This is a flowchart of the decoding module in this invention.

[0020] Figure 4 This is a comparison of the real infrared target image, the predicted image of the infrared small target algorithm based on the coupling of visual image and frequency analysis, and the real infrared target image in the verification experiment of this invention. Detailed Implementation

[0021] In this embodiment, an infrared small target detection method based on the coupling of visual image and frequency analysis 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 is a preprocessed infrared image, and K represents the total number of infrared images; let The real infrared target image is ;in, W and H These represent the width and height of each infrared image, respectively, with 1 representing the number of channels in each infrared image; and for... Perform data augmentation to obtain the first k Zhang enhanced infrared image Thus, the enhanced infrared image dataset is obtained. In this embodiment, the width of the preprocessed infrared image and the enhanced infrared image are... W =265, H =256.

[0022] Step 2: Construct an infrared small target detection network, including: an encoder module based on residual blocks, a space-frequency coupled topology module, a decoding module, and a prediction module, and perform... Processing is performed to obtain the first... k Zhang Hongguang Significance Prediction Chart ; Step 2.1, the Encoder module includes... M The structure consists of layers, each layer comprising a residual block and a max-pooling layer, and these layers are sequentially processed. After processing, the processed infrared coded feature sequence is normalized in the spatial dimension, and then the normalized infrared coded feature sequence is concatenated in the channel dimension to form the first... k Infrared global features In this embodiment, M =4.

[0023] Step 2.1.1 Passing through in sequence M After processing the layer structure, the first layer is obtained. k Infrared coded feature sequence ,in, , No. i The infrared coding features output by the layer structure are obtained by equation (1). Indicates the first i The number of channels for layer infrared coded features; Represents a constant; in this embodiment, r=32, which is the number of output channels of the first-layer encoder block, therefore we can deduce Number of channels , Number of channels , Number of channels 128, Number of channels 256.

[0024] (1) In equation (1), This represents a 2D convolution operation with a 3×3 kernel, a stride of 1, and zero padding. This represents a 2D convolution operation with a 1×1 kernel, a stride of 1, and zero padding. For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation with a step size of 2.

[0025] Step 2.1.2: Use bilinear interpolation to perform spatial interpolation. After normalization, we obtain the first... k A normalized infrared coded feature sequence ,in, express The Middle i A normalized infrared coded feature; in this embodiment... .

[0026] Step 2.1.3, on the channel dimension After concatenating each infrared coded feature in the image, we obtain the first... k Infrared global features Therefore Number of channels The height and width are 64.

[0027] Step 2.2: The space-frequency coupled topology module consists of a visual interactive graph module and a dual-branch spectrum-enhanced frequency control feedforward network, as shown below. Figure 2 As shown, this module is used to... Processing is performed to obtain the first... k Infrared Enhancement Feature Sequence ; Step 2.2.1: The visual interaction diagram module will... Evenly divided into P A set of blocks of the same size ,in, Indicates the first i Each block. Next, calculate... and The Euclidean distance between other blocks in the middle is selected, and then... The closest Each block is connected to form a structure and The edges between the blocks, thus obtaining The edge connection relationships between all blocks in the array are used to generate the first... A binary adjacency matrix If there is an edge connecting the two blocks, then... The corresponding row and column element is set to "1"; otherwise, it is set to "0", thus determining the order based on the given information. and Construct the first k Individual diagram structure Finally, the maximum relative graph convolution operator is used to... Perform feature encoding to obtain the first k Interaction graph features In this embodiment, the following settings are provided: , .

[0028] Step 2.2.2: The frequency-controlled feedforward network first divides the channel along the channel dimension... All are divided into the first infrared interactive image sub-features Second infrared interaction map features and respectively using a core size of and Grouped convolution pairs and Spatial local refinement is performed, resulting in the first infrared refined interactive map sub-feature. Second infrared refined interactive image features Next, using equations (2) and (3) to... and The first spectral modulation coefficient is obtained through processing. Second spectrum modulation coefficient : (2) (3) In equations (2) and (3), rfft2 represents the Fast Fourier Transform; This indicates that the amplitude is taken, and MLP(·) represents a multilayer perceptron; This indicates channel attention gating. This represents the spectrum gating to be learned; in this embodiment, It is 3. It is 5.

[0029] Next, the first decoupling reconstruction feature is obtained using equations (4) and (5). Second decoupling reconstruction features ; (4) (5) In equations (4) and (5), This represents spatial attention gating, and rfft2 represents the Fast Fourier Transform. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. e It is a natural constant.

[0030] Finally, using equations (6) and (7), we obtain the first... k feedforward features : (6) (7) In equations (6) and (7), Indicates the first k One aggregation feature, This indicates a splicing operation along the channel dimension. It is the ReLU activation function; ECA This indicates the channel attention mechanism.

[0031] Step 2.2.3, along the channel dimension Cut into M block, get the first k A set of blocks ,in, Indicates the first i Each block; Step 2.2.4: Use bilinear interpolation to... Length and width with Transform it to be consistent, and obtain the first... k A transformed block set ,in, express The transformed first i Each block; Step 2.2.5: Transform the block set With infrared coded feature sequences Add them according to their indices to get the first... k Infrared enhancement features ,in, ,express The Middle i One infrared enhancement feature.

[0032] Step 2.3, Decoding Module, Decoding Process, as follows: Figure 3 As shown. The decoding module... Processing is performed to obtain the first... k Infrared decoding feature sequence ;in, Indicates the first i One infrared decoding feature; when When =1, the first equation is obtained using equation (8). i- 1 infrared decoding feature : (8) In equation (8), This represents a 2D convolution operation with a 3×3 kernel, a stride of 1, and zero padding. This represents a 2D convolution operation with a 1×1 kernel, a stride of 1, and zero padding. For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation with a step size of 2.

[0033] when =1,2,…, M At that time, the first i Infrared enhancement features Will with the i- 1 infrared decoding feature By merging, we obtain the first... i A fused infrared feature And using equation (8) to obtain the first i Infrared decoding features Thus, the first k Infrared decoding feature sequence ; (9) And for the first Infrared decoding features Perform bilinear interpolation to obtain the first... Infrared interpolation features Therefore, using equations (10), (11), (12), and (13), we obtain : (10) (11) (12) (13) In equations (10)-(13), express n × n Two-dimensional convolution operation, ECA This represents channel attention, and rfft2 represents the Fast Fourier Transform. GAP This indicates the average pooling operation. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. It is a natural constant. This indicates taking its amplitude, and irfft2 represents the Fast Fourier Transform. Indicates by The generated weight map, Indicates by The generated weight map, express Activation function Indicates the first k A final weighted graph, express Activation function.

[0034] Step 2.4, the prediction module... Processing is performed to obtain the first... k Infrared splicing saliency prediction map ; Step 2.4.1: Use equation (14) to obtain the first... i Infrared saliency prediction map Thus, the first k Infrared saliency prediction map ; (14) In equation (14), This is a bilinear interpolation operation. This is the Sigmoid activation function.

[0035] Step 2.4.2: Use equation (15) to obtain the first... k Infrared splicing saliency prediction map : (15) In equation (14), This indicates splicing along the channel dimension.

[0036] Step 3: Construct the total loss of the infrared small target detection network using equation (16). ; (16) In equation (15), Indicates 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; Indicates the first k One infrared saliency prediction map; in this embodiment, and Set to 1.

[0037] 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.

[0038] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the methods described above, and the processor is configured to execute the program stored in the memory.

[0039] 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.

[0040] To verify the effectiveness of this invention, this embodiment applies the method to an infrared small target detection task. In this embodiment, four challenging infrared small target scene images are selected for visualization analysis, and the results are as follows. Figure 4 As shown. In Figure 4 In the diagram, each row sequentially displays: a real infrared image, a predicted image generated by the proposed method, and a real infrared target image. To clearly show the details of small targets, the target area in each of the three images is magnified and placed in the lower right corner. The target location in the original image is marked with a blue circle, and the magnified area is indicated by a blue border. The two are connected by a blue guide line for easy observation and comparison. From the magnified area, it can be observed that the target predicted by the model of this invention highly matches the real annotation in both position and outline shape. This visualization result intuitively demonstrates the accuracy and reliability of the proposed method for detecting small infrared targets in complex scenes.

[0041] To verify the superiority of this invention, this embodiment performs a small target detection task both domestically and internationally, and compares its performance with that of current mainstream infrared small target detection algorithms on the NUDT-SIRST dataset. Table 1 shows the detection accuracy of current mainstream infrared small target detection algorithms on the NUDT-SIRST dataset and compares it with the method of this invention.

[0042] Table 1. Performance comparison with other algorithms in infrared small target detection accuracy.

[0043] To objectively evaluate the performance of the method of this invention, comparative experiments were conducted on the same test set with current mainstream infrared small target detection algorithms, including ACM, DNANet, UIUNet, HCFNet, SCTransNet, and GCLNet. 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. A smaller value indicates a lower false positive rate and better performance. As can be seen from the results in Table 1, the method of this invention performs well in terms of IoU, nIoU, and... Both achieved the highest values ​​in F1 and maintained the minimum. This fully demonstrates that the method of the present invention has superior overall performance and higher detection reliability in infrared small target detection tasks.

Claims

1. A method for detecting small infrared targets based on the coupling of visual image and frequency analysis, characterized in that, Includes the following steps: Step 1: Obtain the preprocessed infrared image dataset ;in, Indicates the first k Zhang is a preprocessed infrared image, and K represents the total number of infrared images; let The real infrared target image is ;in, W and H These represent the width and height of each infrared image, respectively, and 1 represents the number of channels in each infrared image; 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: an encoder module based on residual blocks, a space-frequency coupled topology module, a decoding module, and a prediction module, and perform... Processing is performed to obtain the first... k Zhang Hongguang Significance Prediction Chart ; Step 2.1, the Encoder module includes... M The structure consists of layers, each layer comprising a residual block and a max-pooling layer, and these layers are sequentially processed. After processing, the processed infrared coded feature sequence is normalized in the spatial dimension, and then the normalized infrared coded feature sequence is concatenated in the channel dimension to form the first... k Infrared global features ; Step 2.2: The space-frequency coupled topology module consists of a visual interactive graph module and a dual-branch spectrum-enhanced frequency control feedforward network, and... Processing is performed to obtain the first... k Infrared Enhancement Feature Sequence ; Step 2.3, Decoding module Processing is performed to obtain the first... k Infrared decoding feature sequence ;in, Indicates the first i One infrared decoding feature; Step 2.4, the prediction module... Processing is performed to obtain the first... k Infrared splicing saliency prediction map ; Step 3: Construct the total loss of the infrared small target detection network using equation (1). ; (1) In equation (1), Indicates 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; Indicates the first k One infrared saliency prediction map; 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 infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 1, characterized in that, 2.1 includes: Step 2.1.1 Passing through in sequence M After processing the layer structure, the first layer is obtained. k Infrared coded feature sequence ,in, Indicates the first i Infrared encoded features output by the layer structure Indicates the first i The number of channels for layer infrared coded features; Represents a constant; Step 2.1.2: Use bilinear interpolation to perform spatial interpolation. After normalization, we obtain the first... k A normalized infrared coded feature sequence ,in, express The Middle i A normalized infrared coded feature; Step 2.1.3, on the channel dimension After concatenating each infrared coded feature in the image, we obtain the first... k Infrared global features ,in, The number of channels for the global infrared feature.

3. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 2, characterized in that: The first encoding module in section 2.1.1 i The layered structure is obtained using equation (2). : (2)。 4. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 1, characterized in that, Step 2 includes: Step 2.2.1, Visual Interaction Graph Module Processing is performed to obtain the first... k Infrared interactive graph features ; Step 2.2.2: The frequency-controlled feedforward network uses Fourier transform and gated spectral attention pairs... Processing is performed to obtain the first... k Infrared feedforward features ; Step 2.2.3, along the channel dimension Cut into M block, get the first k A set of blocks ,in, Indicates the first i Each block; Step 2.2.4: Use bilinear interpolation to... Length and width with Transform it to be consistent, and obtain the first... k A transformed block set ,in, express The transformed first i Each block; Step 2.2.5: Transform the block set With infrared coded feature sequences Add them according to their indices to get the first... k Infrared enhancement features ,in, ,express The Middle i One infrared enhancement feature.

5. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 4, characterized in that, Step 2.2.1 includes: Step 2.2.1.1, will Evenly divided into P A set of blocks of the same size ,in, Indicates the first i One block; Step 2.2.1.2, Calculation and The Euclidean distance between other blocks in the middle is selected, and then... The closest Each block is connected to form a structure and The edges between the blocks, thus obtaining The edge connection relationships between all blocks in the array are used to generate the first... A binary adjacency matrix If there is an edge connecting the two blocks, then... The corresponding row and column element is set to "1"; otherwise, it is set to "0", thus determining the order based on the given information. and Construct the first k Individual diagram structure ; Step 2.2.1.3: Use the maximum relative graph convolution operator to pair Perform feature encoding to obtain the first k Interaction graph features .

6. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 4, characterized in that, Step 2.2.2 includes: Step 2.2.2.1, along the channel dimension All are divided into the first infrared interactive image sub-features Second infrared interaction map features and respectively using a core size of and Grouped convolution pairs and Spatial local refinement is performed, resulting in the first infrared refined interactive map sub-feature. Second infrared refined interactive image features ; Step 2.2.2.2: Using equations (3) and (4) to... and The first spectral modulation coefficient is obtained through processing. Second spectrum modulation coefficient : (3) (4) In equations (3) and (4), rfft2 represents the Fast Fourier Transform; This indicates that the amplitude is taken, and MLP(·) represents a multilayer perceptron; This indicates channel attention gating. This represents the spectrum gating to be learned; Step 2.2.2.3: Obtain the first decoupling reconstruction feature using equations (5) and (6). Second decoupling reconstruction features ; (5) (6) In equations (5) and (6), This represents spatial attention gating, and rfft2 represents the Fast Fourier Transform. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. e It is a natural constant; Step 2.2.2.4: Use equations (7) and (8) to obtain the first... k feedforward features : (7) (8) In equations (7) and (8), Indicates the first k One aggregation feature, This indicates a splicing operation along the channel dimension. is the ReLU activation function; ECA represents the channel attention mechanism.

7. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 1, characterized in that, Step 2.3 includes: when When =1, the first equation is obtained using equation (9). i- 1 infrared decoding feature : (9) In equation (9), Indicates that the convolution kernel is n × n Two-dimensional convolution operation, For batch normalization operations, LR The LeakyReLU activation function is used. This is a max pooling operation; when When =1,2,…,M, the first... i Infrared enhancement features Will with the i- 1 infrared decoding feature By merging, we obtain the first... i A fused infrared feature And using equation (10) to obtain the first i Infrared decoding features Thus, the first k Infrared decoding feature sequence ; (10)。 8. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 7, characterized in that, Step 2.3 is for the first Infrared decoding features Perform bilinear interpolation to obtain the first... Infrared interpolation features Therefore, using equations (11), (12), (13), and (14), we obtain : (11) (12) (13) (14) In equations (11)-(14), express n × n Two-dimensional convolution operation, ECA This represents channel attention, and rfft2 represents the Fast Fourier Transform. GAP This indicates the average pooling operation. The imaginary unit is represented by ∠·, which indicates the extraction of its phase. It is a natural constant. This indicates taking its amplitude, and irfft2 represents the Fast Fourier Transform. Indicates by The generated weight map, Indicates by The generated weight map, express Activation function Indicates the first k A final weighted graph, express Activation function.

9. The infrared small target detection method based on the coupling of visual image and frequency analysis according to claim 1, characterized in that, Step 4 includes: Step 2.4.1: Use equation (15) to obtain the first... i Infrared saliency prediction map Thus, the first k Infrared saliency prediction map ; (15) In equation (17), This is a bilinear interpolation operation. Use the Sigmoid activation function; Step 2.4.2: Using equation (16) to obtain the first... k Infrared splicing saliency prediction map : (16) In equation (18), This indicates splicing along the channel dimension.

10. 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 1-9, the processor being configured to execute the program stored in the memory.

Images