A single infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing

Abstract

This invention discloses a single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing, relating to the fields of infrared image processing and target detection technology. The method includes frequency domain adaptive enhancement and spatial detail focusing steps, implemented based on a U-Net-like network: an encoding module extracts multi-level features and enhances the saliency of small targets through frequency domain transformation, frequency masking, and channel interaction; a cross-scale feature interaction module achieves multi-scale information complementarity; and a decoding and reconstruction module focuses on small target details through channel attention, local refinement operators, and gated residual modulation. This invention overcomes the problems of poor background suppression in traditional methods and feature degradation in deep learning methods, achieving high-precision, low-false-alarm detection of infrared small targets in complex backgrounds, and is suitable for scenarios such as infrared search and tracking, and early warning detection.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing and target detection technology, and in particular to a method for detecting small targets in a single infrared image. Specifically, it is a single-frame infrared small target detection method that combines frequency domain saliency enhancement and spatial detail focusing, which can be applied to scenarios such as infrared search and tracking, early warning detection, unmanned system perception, land, sea and air monitoring, traffic management and earth observation. Background Technology

[0002] The core task of infrared small target detection is to identify weak targets occupying only a few pixels in complex backgrounds, which plays a crucial role in many fields such as land, sea and air monitoring, traffic management and earth observation. However, infrared imaging is susceptible to factors such as distance attenuation, optical system bandwidth limitations and random noise interference, resulting in infrared small targets generally exhibiting the characteristics of being "weak, few and messy": the target signal is weak, local texture is almost missing, and background clutter is strong and varied.

[0003] Traditional infrared small target detection methods include filtering methods, background modeling-based algorithms, and local contrast mechanisms. These methods typically rely on hand-designed features and background models, and their effectiveness in background modeling and noise suppression is unsatisfactory in complex backgrounds. In recent years, deep learning technology, especially convolutional neural networks, has made significant progress in this field, enabling automatic learning of multi-level features in images and achieving relatively accurate detection. However, several bottlenecks remain, including: the spatial information loss caused by downsampling further weakens the ability of deep semantics to model extremely small targets; few upsampling modules have been optimized or modified specifically for infrared small target detection tasks; and insufficient cross-scale feature fusion makes it difficult to effectively suppress false bright spots in complex backgrounds, resulting in insufficient detection stability. Summary of the Invention

[0004] The purpose of this invention is to overcome the technical defects of the existing technology and provide a single-frame infrared small target detection method that combines frequency domain saliency enhancement and spatial detail focusing. It enhances the saliency response of small targets in the frequency domain and adaptively focuses on key details in the spatial domain, thereby achieving high-precision and low false alarm detection of infrared small targets in complex backgrounds. This solves the problems of poor background modeling and noise suppression in traditional methods, spatial information loss, feature degradation, and false bright spot detection in deep learning methods.

[0005] This invention discloses a single-image infrared small target detection method that combines frequency domain saliency enhancement and spatial detail focusing. The method includes a frequency domain adaptive enhancement step and a spatial detail focusing step. Infrared small target detection is achieved through an encoding module, a cross-scale feature interaction module, and a decoding and reconstruction module of a U-Net-like network. The encoding module performs multi-level feature extraction on the input infrared image and introduces frequency domain enhancement processing. The cross-scale feature interaction module performs joint modeling of spatial and channel dimensions of features at different scales. The decoding and reconstruction module obtains the detection result through stepwise upsampling, feature fusion, local detail refinement, and gated modulation mechanisms.

[0006] Preferably, the frequency domain adaptive enhancement step includes: Step 1) Infrared Image Input and Preprocessing: Acquire a single infrared image as input and feed it into the feature extraction network to obtain the initial feature map. Where C is the number of feature channels, H is the feature map height, and W is the feature map width, the features are denoised using depthwise separable convolution, and the formula is as follows: ,in This represents the channel-wise depthwise convolution operator. These are the denoising weighting coefficients; Step 2) Frequency Domain Information Processing: The denoised features are mapped to the frequency domain using Discrete Cosine Transform (DCT), as shown in the formula below. ,in The discrete cosine transform operator is represented; a frequency mask M is constructed to suppress low-frequency responses, and a spatial saliency guide map is generated by mapping back to the spatial domain through the inverse discrete cosine transform. The formula is as follows: ,in This represents the inverse discrete cosine transform operator. This represents the element-wise multiplication operation; Step 3) Channel Information Modeling: Adaptive pooling is performed on the spatial saliency guidance graph to aggregate global context information. Channel weight vectors are generated through nonlinear mapping, as shown in the formula. ,in This represents a combination of global average pooling and max pooling operations. For channel mapping functions, Use the Sigmoid activation function; Step 4) Feature Fusion Enhancement: The channel weight vectors are fused with the spatial saliency guide map channel by channel, using the following formula: The frequency domain enhanced output features are obtained. .

[0007] Preferably, the spatial detail focusing step includes: Step 1) Feature Fusion: Obtain the decoding stage features and the corresponding scale jump connection features. The decoding stage features are upsampled to restore the spatial resolution consistent with the jump connection features. The jump connection features are modeled through the channel interaction attention mechanism CCA and then concatenated with the upsampled decoding features in the channel dimension to form fused features. Step 2) Feature Dimensionality Reduction and Adaptive Kernel Construction: This involves processing the concatenated fused features... Dimensionality reduction is achieved through 1x1 convolution, followed by the construction of a spatially adaptive kernel using a dynamic convolution kernel generator; Step 3) Local thinning modeling: Introduce local thinning operators in the low-dimensional feature space. Modeling the local structural information of the target neighborhood, and restoring the channel dimension through 1x1 convolution, the formula is as follows: ,in and These represent the convolution weights during the channel compression and channel restoration stages, respectively. Indicates the convolution operation; Step 4) Gated Residual Modulation: Adaptive modulation of the refinement branch is performed through a learnable gated residual mechanism, as shown in the formula. ,in These are learnable scalar parameters, initialized to 0 during the initial training phase.

[0008] Preferably, the basic modules in the U-Net network can be improved based on ResNetBlock.

[0009] Preferably, the cross-scale feature interaction module achieves information complementarity of features at different scales through the Spatial-Channel Cross Transformer Block (SCTB).

[0010] Preferably, the upsampling operation is implemented using bilinear interpolation (BI).

[0011] Preferably, the frequency mask M is constructed based on the characteristics of infrared small targets corresponding to mid-to-high frequency components and background texture energy concentrated in the low-frequency region, in order to specifically suppress low-frequency response.

[0012] Preferably, the Channel Interaction Attention (CCA) mechanism is used to enhance the effective channel response related to small infrared targets and suppress redundant background information.

[0013] Preferably, the local refinement operator The edge and structural features of small infrared targets are enhanced by weighting the response of small-scale neighborhoods, while suppressing large-scale background interference.

[0014] The present invention provides an infrared small target detection system using the method described above, characterized in that it comprises an image input module, a feature extraction module, a frequency domain enhancement module, a cross-scale feature interaction module, a spatial focusing module, and a detection result output module connected in sequence. The system comprises the following components: an image input module output connected to a feature extraction module input, transmitting a single infrared image to the feature extraction module; a feature extraction module output connected to a frequency domain enhancement module input, extracting an initial feature map based on an improved ResNetBlock and transmitting it to the frequency domain enhancement module; a frequency domain enhancement module output connected to a cross-scale feature interaction module input, performing the frequency domain adaptive enhancement step and transmitting the enhanced features to the cross-scale feature interaction module; a cross-scale feature interaction module output connected to a spatial focusing module input, performing multi-scale feature information complementarity through a spatial-channel cross-transformer block (SCTB) and transmitting the information to the spatial focusing module; and a spatial focusing module output connected to a detection result output module input, performing the spatial detail focusing step to refine and enhance the features before transmitting them to the detection result output module, which then outputs the final infrared small target detection result.

[0015] Compared with the prior art, the technical solution of this application has the following beneficial effects: Frequency-domain adaptive enhancement uses discrete cosine transform and frequency masking to specifically highlight the mid-to-high frequency components corresponding to small targets, suppress low-frequency background interference, and improve the saliency of small targets. The channel interaction mechanism models the differences in contributions of different channels, strengthens the effective feature response, and improves the feature representation capability. Spatial detail focusing accurately captures the edge and structural features of small targets through channel interaction attention mechanism, local refinement operator, and gated residual modulation, suppressing background noise and false bright spots. The U-Net-like network combined with cross-scale feature interaction achieves multi-scale information complementarity, alleviates the spatial information loss caused by downsampling, and improves detection stability. The overall method does not require complex preprocessing, has high detection accuracy and low false alarm rate, and is suitable for infrared small target detection in complex backgrounds, with a wide range of application scenarios. Attached Figure Description

[0016] Figure 1 This is a block diagram illustrating the principle of the frequency domain adaptive enhancement method of this invention.

[0017] Figure 2 This is a schematic diagram of the principle of the spatial detail focusing method of the present invention.

[0018] Figure 3 This is a block diagram illustrating the principle of the single-frame infrared small target detection network of the present invention, which combines frequency domain saliency enhancement and spatial detail focusing. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Invention content / principle: A frequency domain adaptive enhancement method, comprising the following steps: Infrared Image Input and Preprocessing: This invention employs a pre-processing technique for infrared images to preserve valuable information. Given an input feature map... First, the features are denoised using depthwise separable convolution to reduce the interference of random noise on the frequency domain analysis. This process can be represented as follows: ; in This represents the channel-wise depthwise convolution operator. These are the denoising weight coefficients. This operation enhances feature stability while preserving structural information.

[0021] Frequency domain information processing: This invention employs a novel technical approach, mapping features in infrared images to the frequency domain for analysis. ; Considering that infrared small targets mainly correspond to mid-to-high frequency components, while background texture energy is concentrated in the low-frequency region, a frequency mask M is constructed to suppress the low-frequency response. The inverse discrete cosine transform is used to map the frequency domain features back to the spatial domain, generating a spatial saliency guidance map. ; Channel Information Modeling: To further model the differences in contributions of different feature channels to small targets at the channel dimension, this invention introduces a channel interaction mechanism based on frequency domain enhanced features. Specifically, adaptive pooling is performed to aggregate global context information, and channel weight vectors are generated through nonlinear mapping.

[0022] ; in This represents a combination of global average pooling and max pooling operations. For channel mapping functions, This is the Sigmoid activation function.

[0023] Feature fusion enhancement: The channel weights are fused with the features of the initially processed infrared image to enhance the image. ; A spatial detail focusing method, comprising the following: Feature fusion: Features from the decoding stage and corresponding scaled skip-connected features are obtained, wherein the decoding stage features are upsampled to a spatial resolution consistent with the skip-connected features; the skip-connected features are modeled using a channel interaction attention (CCA) mechanism to enhance the effective channel response related to small infrared targets and suppress redundant background information; subsequently, the upsampled decoding features and the skip-connected features modeled by channel interaction are concatenated along the channel dimension to form a fused feature for subsequent spatial detail focusing.

[0024] Feature Dimensionality Reduction and Adaptive Kernel Construction: Given Upsampled Concatenated Features First, dimensionality reduction is achieved through 1x1 convolution, and then a spatially adaptive kernel is constructed through a dynamic convolution kernel generator to achieve differentiated modeling of feature responses at different spatial locations.

[0025] Local thinning modeling: Introducing local thinning operators in a low-dimensional feature space The model aims to model the local structural information within the target's neighborhood and recover the channel dimension through 1x1 convolution. By weighting the response of the small-scale neighborhood, the edge and structural features of infrared small targets are enhanced, while suppressing interference from the large-scale background.

[0026] ; in, and These represent the convolution weights during the channel compression and channel recovery stages, respectively.

[0027] Gated residual modulation: To avoid excessive amplification of background noise by local enhancement, a learnable gated residual mechanism is introduced to adaptively modulate the refinement branch. Its output is defined as: ; in These are learnable scalar parameters, initialized to 0 in the early stages of training to ensure approximately identical mappings between modules, thus guaranteeing the stability of the decoding process; as training progresses, Adaptive adjustment allows local refinement branches to gradually participate in feature reconstruction, enhancing the expression of small target edges and structures.

[0028] Combining frequency domain saliency enhancement and spatial detail focusing, this invention also proposes a high-efficiency detection network for infrared small target detection, specifically including the following: Basic Module: The basic module used in this invention is based on an improvement of the commonly used ResNet Block.

[0029] Network Structure: This invention employs a U-Net-like network, whose overall structure includes an encoding module, a cross-scale feature interaction module, and a decoding and reconstruction module. The encoding module extracts multi-level features from the input infrared image and introduces frequency domain enhancement processing into features at different scales to highlight high-frequency salient information related to small targets. The cross-scale feature interaction module jointly models features at different scales in both spatial and channel dimensions, achieving information complementarity between multi-scale features. The decoding and reconstruction module refines the small target region through progressive upsampling and feature fusion, combined with local detail refinement and gated modulation mechanisms, thereby obtaining the final target detection result.

[0030] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] Example 1: Implementation of Frequency Domain Adaptive Enhancement Module like Figure 1 This is a block diagram illustrating the principle of the frequency domain adaptive enhancement method of the present invention. This example includes the following: (1) Infrared image input and preprocessing: Acquire a single infrared image and input it into the feature extraction network to obtain the initial feature map. Through the channel-wise depth convolution operator Perform noise reduction processing, such as setting noise reduction weight coefficients. According to the formula Calculate the denoised feature map This operation enhances feature stability while preserving structural information; (2) Frequency domain saliency enhancement processing: using the discrete cosine transform operator right Perform frequency domain transformation to obtain Based on the frequency distribution characteristics of infrared small targets and the background, a frequency mask M is constructed. For example, a suppression weight of 0.2 is assigned to the low-frequency region, and a retention weight of 1.0 is assigned to the mid-to-high-frequency region. Mapping frequency domain features back to the spatial domain generates a spatial saliency guide map. (3) Channel information modeling: for Perform a combination of global average pooling and max pooling operations. The obtained feature input channel mapping function (Implemented using a 2-layer fully connected network), and then activated by the Sigmoid function. Generate channel weight vector ; (4) Feature fusion enhancement: and Perform channel-by-channel multiplication operation , Obtain frequency domain enhanced output features This provides high-quality feature input for subsequent spatial detail focusing.

[0032] Example 2: Implementation of the Spatial Detail Focusing Module like Figure 2 This is a schematic diagram of the spatial detail focusing method of the present invention. The example includes the following: (1) Feature fusion: Obtain the features of the decoding stage and the corresponding scale of the skip connection features. Use bilinear interpolation (BI) to upsample the features of the decoding stage so that their spatial resolution is consistent with that of the skip connection features. Input the skip connection features into the channel interaction attention mechanism (CCA) module to enhance the channel response of small targets and suppress redundant background information. Then, concatenate the upsampled decoding features with the skip connection features modeled by CCA in the channel dimension to form the fused features. ; (2) Feature dimensionality reduction and adaptive kernel construction: For example, the number of channels of the fused feature U is reduced from 256 to 64 by 1X1 convolution, and a 3×3 spatial adaptive kernel is constructed by a dynamic convolution kernel generator (composed of 3 layers of convolutional network) to realize differentiated modeling of features at different spatial locations; (3) Local thinning modeling: Introducing local thinning operators in the low-dimensional feature space (LSConv), setting the neighborhood receptive field to 3×3, for ( Local structure modeling is performed on the 1x1 dimensionality-reduced convolution weights, and then... (1x1 up-dimensional convolution weights) restore the number of channels to 256, resulting in ; (4) Gated residual modulation: such as initialization of learnable scalar parameters According to the formula Calculate the output during training. Adaptive adjustment gradually enhances the contribution of locally refined branches, improving the expression of small target edges and structures.

[0033] Example 3: Overall Implementation of the Detection Network like Figure 3 This is a block diagram illustrating the principle of the single-frame infrared small target detection network combining frequency domain saliency enhancement and spatial detail focusing according to the present invention. This example includes the following: (1) Basic network module: The basic module based on ResNetBlock is adopted, and the frequency domain adaptive enhancement module (AFEM) is connected after the output of each residual block. (2) Encoding module: For example, the input infrared image is processed through four encoding stages for feature extraction. Each encoding stage includes two improved ResNetBlock and one max pooling operation. The output features of each stage are processed by AFEM to highlight the high-frequency salient information of small targets. (3) Cross-scale feature interaction module: The spatial-channel cross-transformer block (SCTB) is used to jointly model the spatial and channel dimensions of the four scale features output by the encoding module, so as to achieve multi-scale information complementarity; (4) Decoding and reconstruction module: It includes 4 decoding stages. Each decoding stage realizes feature reconstruction through bilinear interpolation upsampling, feature concatenation and 3 convolutional layers. After concatenation, a spatial detail focusing module (including CCA and GLRC) is introduced to refine the small target area. (5) Output layer: The final decoded features are mapped to a single-channel detection map through 1x1 convolution, and the infrared small target detection results are obtained by threshold segmentation.

[0034] Finally, it should be noted that the present invention is not limited to the above embodiments, and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting single infrared small target by combining frequency domain saliency enhancement and spatial detail focusing, characterized in that, The method includes a frequency domain adaptive enhancement step and a spatial detail focusing step. Infrared small target detection is achieved through a U-Net-like network encoding module, a cross-scale feature interaction module, and a decoding and reconstruction module. The encoding module performs multi-level feature extraction on the input infrared image and introduces frequency domain enhancement processing. The cross-scale feature interaction module performs joint modeling of spatial and channel dimensions of features at different scales. The decoding and reconstruction module obtains the detection result through stepwise upsampling, feature fusion, local detail refinement, and gated modulation mechanisms.

2. The single-frame infrared dim small target detection method of claim 1, wherein, The frequency domain adaptive enhancement step includes: Step 1) infrared image input and preprocessing: a single infrared image is obtained as input, and is input into a feature extraction network to obtain an initial feature map wherein C is the number of feature channels, H is the height of the feature map, and W is the width of the feature map, the features are denoised by depth separable convolution, and the formula is , wherein represents a channel-by-channel depth convolution operator, is a denoising weight coefficient; Step 2) Frequency domain information processing: the denoised features are mapped to the frequency domain space using a discrete cosine transform, with the formula , where represents the discrete cosine transform operator; a frequency mask M is constructed to suppress low-frequency responses, and the spatial saliency guide map is generated by mapping back to the spatial domain using an inverse discrete cosine transform, with the formula , where represents the inverse discrete cosine transform operator, represents an element-wise multiplication operation; Step 3) Channel Information Modeling: Adaptive pooling is performed on the spatial saliency guidance graph to aggregate global context information. Channel weight vectors are generated through nonlinear mapping, as shown in the formula. ,in This represents a combination of global average pooling and max pooling operations. For channel mapping functions, Use the Sigmoid activation function; Step 4) Feature Fusion Enhancement: The channel weight vectors are fused with the spatial saliency guide map channel by channel, using the following formula: The frequency domain enhanced output features are obtained. .

3. A single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 1 or 2, characterized in that, The spatial detail focusing steps include: Step 1) Feature Fusion: Obtain the decoding stage features and the corresponding scale jump connection features. The decoding stage features are upsampled to restore the spatial resolution consistent with the jump connection features. The jump connection features are modeled through the channel interaction attention mechanism CCA and then concatenated with the upsampled decoding features in the channel dimension to form fused features. Step 2) Feature Dimensionality Reduction and Adaptive Kernel Construction: This involves processing the concatenated fused features... Dimensionality reduction is achieved through 1x1 convolution, followed by the construction of a spatially adaptive kernel using a dynamic convolution kernel generator; Step 3) Local thinning modeling: Introduce local thinning operators in the low-dimensional feature space. Modeling the local structural information of the target neighborhood, the channel dimension is recovered through 1x1 convolution, as shown in the formula. ,in and These represent the convolution weights during the channel compression and channel restoration stages, respectively. Indicates the convolution operation; Step 4) Gated Residual Modulation: Adaptive modulation of the refinement branch is performed through a learnable gated residual mechanism, as shown in the formula. ,in These are learnable scalar parameters, initialized to 0 during the initial training phase.

4. The single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 1, characterized in that, The basic modules in the U-Net network can be improved based on ResNetBlock.

5. The single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 1, characterized in that, The cross-scale feature interaction module achieves information complementarity of features at different scales through the Spatial-Channel Cross Transformer Block (SCTB).

6. The single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 3, characterized in that, The upsampling operation is implemented using bilinear interpolation (BI).

7. A single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 2, characterized in that, The frequency mask M is constructed based on the characteristics of infrared small targets corresponding to mid-to-high frequency components and background texture energy concentrated in the low-frequency region, and is used to specifically suppress low-frequency response.

8. The single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 3, characterized in that, The Channel Interactive Attention (CCA) mechanism is used to enhance the effective channel response related to small infrared targets and suppress redundant background information.

9. A single-frame infrared small target detection method combining frequency domain saliency enhancement and spatial detail focusing according to claim 3, characterized in that, The local refinement operator The edge and structural features of small infrared targets are enhanced by weighting the response of small-scale neighborhoods, while suppressing large-scale background interference.

10. An infrared small target detection system applying the method of any one of claims 1-9, characterized in that, It includes an image input module, a feature extraction module, a frequency domain enhancement module, a cross-scale feature interaction module, a spatial focusing module, and a detection result output module connected in sequence; The image input module's output is connected to the feature extraction module's input, transmitting a single infrared image to the feature extraction module. The feature extraction module's output is connected to the frequency domain enhancement module's input, whereby the feature extraction module extracts an initial feature map based on an improved ResNetBlock and transmits it to the frequency domain enhancement module. The frequency domain enhancement module's output is connected to the cross-scale feature interaction module's input, whereby the frequency domain enhancement module performs the frequency domain adaptive enhancement step described in claim 2, transmitting the enhanced features to the cross-scale feature interaction module. The cross-scale feature interaction module's output is connected to the spatial focusing module's input, whereby the cross-scale feature interaction module achieves multi-scale feature information complementarity through a spatial-channel cross-transformer block (SCTB) and then transmits it to the spatial focusing module. The spatial focusing module's output is connected to the detection result output module's input, whereby the spatial focusing module performs the spatial detail focusing step described in claim 3, refining the features before transmitting them to the detection result output module, which then outputs the final infrared small target detection result.

Images