Infrared small target segmentation method and device

Abstract

The present application relates to the technical field of image processing, and provides an infrared small target segmentation method and device, the infrared small target segmentation method comprising: acquiring an input image; obtaining a multi-level feature map through a feature extraction model, performing spatial and frequency domain fusion on the extracted features to obtain fused features, dynamically generating a frequency band filter kernel and performing frequency band enhancement to obtain enhanced features; decoding based on the multi-level feature map to obtain decoded features and an initial prediction map, dynamically generating a high-pass filter kernel, using the initial prediction map to guide the enhanced input image to obtain a guided image, performing high-frequency filtering using the high-pass filter kernel and correcting the initial prediction map to obtain a segmentation result. The present application takes into account global and local perception, effectively enhances target saliency and suppresses background clutter, and improves segmentation accuracy.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an infrared small target segmentation method and apparatus. Background Technology

[0002] Infrared small target segmentation is one of the core technologies in infrared detection and search systems, widely used in long-range reconnaissance and early warning, military guidance, and situational awareness. Currently, infrared small target segmentation typically employs deep learning networks based on encoder-decoder structures for feature extraction and pixel-level classification. However, existing technologies still face several challenges in practical applications: on the one hand, conventional feature extraction modules often rely on spatial convolution operations, whose receptive field is limited by the kernel size, making it difficult to effectively capture long-range dependencies and global contextual information, resulting in insufficient modeling capabilities for weak targets in complex backgrounds; on the other hand, in the decoding and prediction stage, existing methods often directly utilize semantic features for upsampling recovery, lacking targeted enhancement mechanisms for high-frequency target signals, making it difficult to effectively distinguish between low-frequency background clutter and high-frequency target signals, easily leading to false alarms or missed detections. Therefore, there is an urgent need for an infrared small target segmentation method that can balance global perception and local details while dynamically enhancing target saliency. Summary of the Invention

[0003] To address the limitations of existing technologies in extracting weak target features under complex backgrounds and insufficient target saliency during the prediction stage, this application proposes an infrared small target segmentation method and apparatus that achieves adaptive frequency band enhancement of target features and dynamic and accurate highlighting of high-frequency signals.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an infrared small target segmentation method, comprising the following steps: Obtain the input image; The input image is processed by a feature extraction model to obtain a multi-level feature map. During the feature extraction process, the extracted features are fused by spatial and frequency domains to obtain fused features. A frequency band filter kernel is dynamically generated, and the fused features are adaptively enhanced using the frequency band filter kernel to obtain enhanced features. Decoding is performed on the multi-level feature maps to obtain decoded features and an initial prediction map; a high-pass filter kernel is dynamically generated, and the input image is enhanced using the initial prediction map to obtain a guide image; the guide image is then subjected to high-frequency filtering using the high-pass filter kernel, and the initial prediction map is corrected based on the high-frequency filtering result to obtain the target segmentation result.

[0005] The above scheme achieves complementary enhancement of local texture and global context by introducing spatial and frequency domain fusion processing in the feature extraction stage and combining it with frequency band filter kernels dynamically generated based on full-text information. At the same time, in the decoding and prediction stage, high-pass filter kernels are dynamically generated using decoding features, and the input image is guided by the initial prediction map to perform high-frequency enhancement and feedback correction, which effectively improves the saliency and segmentation accuracy of weak targets.

[0006] As one implementation method, the dynamically generated band filter kernel, used to enhance the fused features to obtain enhanced features, includes: performing global pooling on the fused features to obtain global context information; predicting band modulation parameters based on the global context information, the band modulation parameters including center frequency parameters, bandwidth parameters, and shape parameters; constructing the band filter kernel based on the band modulation parameters; transforming the fused features to the frequency domain, performing frequency domain modulation using the band filter kernel, inversely transforming it back to the spatial domain, and then superimposing and fusing it with the fused features to obtain the enhanced features.

[0007] As one implementation, constructing the band filter kernel based on the band modulation parameters includes: constructing the band filter kernel for the Kth channel according to the following formula. :

[0008] in, For frequency domain coordinates, The Euclidean distance from the frequency domain coordinates to the center point of the frequency domain is used to characterize the frequency magnitude; The center frequency parameter, The bandwidth parameters are used to control the bandwidth. To control the shape parameters of the attenuation steepness, the band filter kernels of each channel are spliced ​​together along the channel dimension to obtain the final band filter kernel matrix. As one implementation, after obtaining the enhanced features, the method further includes: processing the enhanced features through a channel attention module and a spatial attention module to suppress background features and enhance the target response, thereby obtaining the final output features.

[0009] As one implementation method, the dynamic generation of the high-pass filter kernel, using the initial prediction image to enhance the input image to obtain a guide image, using the high-pass filter kernel to perform high-frequency filtering on the guide image, and correcting the initial prediction image based on the high-frequency filtering result to obtain the target segmentation result, includes: performing global pooling processing on the decoded features to obtain global context information; predicting high-pass modulation parameters based on the global context information, the high-pass modulation parameters including bandwidth parameters and shape parameters; constructing the high-pass filter kernel based on the high-pass modulation parameters; performing element-wise operations on the initial prediction image and the input image to obtain the guide image; converting the guide image to the frequency domain, performing high-frequency filtering using the high-pass filter kernel, and then inversely transforming it back to the spatial domain to obtain high-frequency enhancement features; and performing weighted correction on the initial prediction image based on the high-frequency enhancement features and the Sigmoid activation function to obtain the target segmentation result.

[0010] As one implementation, constructing the high-pass filter kernel based on the high-pass modulation parameters includes: constructing the high-pass filter kernel according to the following formula. :

[0011] in, For frequency domain coordinates, The distance from the frequency domain coordinates to the center point of the frequency domain. This is a bandwidth parameter used to control the high-pass filter core. The bandwidth; The shape parameter is used to control the high-pass filter kernel. The steepness of the decay.

[0012] As one implementation method, the step of weighting and correcting the initial prediction map based on the high-frequency enhancement features and the Sigmoid activation function to obtain the target segmentation result includes: inputting the high-frequency enhancement features into the Sigmoid activation function to obtain a Sigmoid weight map; multiplying the initial prediction map and the Sigmoid weight map element-wise, and superimposing the multiplication result with the initial prediction map to obtain the target segmentation result.

[0013] As one implementation, the step of processing the input image through a feature extraction model to obtain a multi-level feature map includes: at the encoding end, extracting multi-level semantic features of the input image layer by layer through multiple downsampling operations and feature extraction operations; the step of decoding based on the multi-level feature map to obtain decoded features and an initial prediction map includes: at the decoding end, restoring the feature map resolution through multiple upsampling operations, and concatenating and fusing the corresponding level of encoding end features with the decoding end features to generate multi-scale features; and generating the initial prediction map corresponding to each scale based on the multi-scale features.

[0014] As one implementation method, the step of correcting the initial prediction map based on the high-frequency filtering result to obtain the target segmentation result includes: performing channel stitching on the corrected initial prediction map corresponding to each scale, and processing it through convolution operation and activation function to output the final binary classification segmentation result map.

[0015] Furthermore, the present invention also provides an infrared small target segmentation device, comprising: The image acquisition module is used to acquire the input image; The hybrid filtering module is used to process the input image through a feature extraction model to obtain a multi-level feature map. During the feature extraction process, the extracted features are subjected to spatial and frequency domain fusion processing to obtain fused features, and a frequency band filter kernel is dynamically generated. The frequency band filter kernel is used to adaptively enhance the fused features to obtain enhanced features. The dynamic high-frequency enhancement module is used to perform decoding processing based on the multi-level feature map to obtain decoded features and an initial prediction map; dynamically generate a high-pass filter kernel, use the initial prediction map to perform guidance enhancement on the input image to obtain a guidance image, use the high-pass filter kernel to perform high-frequency filtering on the guidance image, and correct the initial prediction map based on the high-frequency filtering result to obtain the target segmentation result.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the limitation of the receptive field of traditional spatial convolution by performing spatial and frequency domain fusion processing during the feature extraction stage and dynamically generating frequency band filtering kernels for adaptive frequency band enhancement. Furthermore, it can adaptively select and enhance effective frequency band components according to the characteristics of the input content, achieving deep fusion and complementarity between local detail texture and global context information, and significantly improving the discriminative power of feature representation in complex backgrounds.

[0017] This invention introduces a dynamic high-frequency enhancement mechanism in the decoding and prediction stage. It dynamically constructs a high-pass filter kernel and uses the initial prediction map to guide the enhancement of the input image before performing high-frequency filtering and feedback correction. This design avoids the limitations of fixed filters that cannot adapt to different scenarios, accurately locating and enhancing the high-frequency response signal of the target while effectively suppressing low-frequency background clutter. By using the Sigmoid activation function to feed back the high-frequency enhancement result to correct the initial prediction map, a closed-loop optimization path of "prediction-enhancement-correction" is formed, thereby significantly improving the segmentation accuracy and robustness of weak targets.

[0018] In this invention, the adaptive frequency band enhancement in the feature extraction stage and the dynamic high-frequency enhancement in the decoding stage work synergistically: the former initially completes the selective enhancement of the target frequency band at the feature level, providing high-quality semantic features for the latter; the latter, based on the former, dynamically constructs a high-pass filter kernel using the decoding features to accurately extract and correct high-frequency signals in the guiding image. Both form a complete spectrum optimization link from feature representation to prediction output, jointly achieving a cascaded improvement in the salience of weak targets in complex backgrounds. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the infrared small target segmentation network architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the hybrid filtering module structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the dynamic high-frequency enhancement module structure according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, 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.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] like Figure 1 As shown, this embodiment provides an infrared small target segmentation method. This method achieves accurate extraction and segmentation of target features in complex backgrounds through a two-stage dynamic frequency domain modulation mechanism. It should be understood that, although... Figure 1 A specific network architecture flow is illustrated, but in other embodiments, any variation that includes the following core processing logic falls within the scope of protection of this invention. The method specifically includes the following steps: Step S100: Obtain the input image.

[0023] Specifically, the input image is an infrared image. This invention does not impose specific limitations on the modality, resolution, or acquisition device of the input image. This step aims to provide the raw data foundation for subsequent feature extraction. In practical applications, it may also include conventional operations such as normalization, resizing, or noise preprocessing of the input image to meet the input requirements of the feature extraction model.

[0024] Step S200: The input image is processed by the feature extraction model to obtain a multi-level feature map. During the feature extraction process, the extracted features are fused by spatial and frequency domain fusion processing to obtain fused features. A frequency band filter kernel is dynamically generated, and the fused features are adaptively enhanced by the frequency band filter kernel to obtain enhanced features.

[0025] Specifically, the feature extraction model is used to abstract the semantic information of the input image layer by layer, generating multi-level feature maps with different spatial resolutions and semantic depths. Addressing the limitations of traditional spatial domain convolution in terms of its limited receptive field and difficulty in capturing global contextual information, this embodiment introduces a fusion processing mechanism between the spatial and frequency domains. The spatial domain processing branch focuses on extracting local detail features such as edges and textures, while the frequency domain processing branch utilizes the global characteristics of frequency domain transformation to capture long-distance dependencies. The fused feature obtained after the fusion possesses both local precision and global generalization power. Furthermore, to overcome the limitation of fixed filters in adapting to varying scenes, this embodiment predicts and generates a frequency band filter kernel in real time based on the global contextual information of the fused feature. This frequency band filter kernel can adaptively select and enhance specific frequency band components that are most effective in representing the target, based on the spectral distribution characteristics of the current image content, while suppressing ineffective background noise frequency bands. This "content-adaptive" frequency band enhancement strategy allows the feature extraction process to dynamically focus on the most discriminative information dimension, significantly improving the robustness of features in complex environments. The mathematical expression of this feature extraction process is as follows:

[0026] in, and These represent the output feature map and the input feature map, respectively. This represents a hybrid filtering operation, namely the overall operator for the fusion of spatial and frequency domains and bandwidth enhancement.

[0027] Step S300: Decode the multi-level feature map to obtain the decoded features and the initial prediction map; dynamically generate a high-pass filter kernel; use the initial prediction map to enhance the input image to obtain a guide image; use the high-pass filter kernel to perform high-frequency filtering on the guide image; and correct the initial prediction map based on the high-frequency filtering result to obtain the target segmentation result.

[0028] Specifically, in the decoding stage, the spatial resolution of the feature map is gradually restored through operations such as upsampling, and the decoded features and corresponding initial prediction map are generated by combining the hierarchical features from the encoding end. Although the initial prediction map contains coarse location information of the target, it is often accompanied by background clutter interference and blurred target boundaries. To address this, this embodiment designs a second-stage dynamic high-frequency modulation mechanism. First, a high-pass filter kernel is dynamically generated based on the global information of the decoded features. The parameters of this high-pass filter kernel change dynamically with the semantic content of the decoded features, aiming to accurately match the high-frequency signal features of the target at the current scale. Second, the input image is enhanced using the initial prediction map. This step essentially uses the initial prediction map as a spatial attention mask or prior probability map to weight or modulate the original input image, thereby significantly suppressing background interference in non-target areas while preserving the original high-frequency details of the potential target region, resulting in a guide image with a higher signal-to-noise ratio. Subsequently, the guide image is subjected to high-frequency filtering using the aforementioned dynamically generated high-pass filter kernel to extract the pure target high-frequency enhancement signal. Finally, the high-frequency filtering result is fed back to the initial prediction map for correction, forming a closed-loop optimization path of "prediction-guidance-enhancement-correction". This mechanism effectively solves the problem that weak targets are easily submerged by low-frequency background during the decoding stage. By introducing high-frequency detail feedback from the original image, it significantly improves the accuracy and completeness of the final target segmentation result.

[0029] The dynamic high-frequency enhancement module operates at the decoding end, dynamically enhancing the target's high-frequency signal based on the input characteristics of the initial prediction map at each scale, thereby increasing the target's salience and suppressing low-frequency background clutter. The steps of dynamic high-frequency enhancement are as follows: [The text then abruptly shifts to a different topic:] The decoding end... Each scale feature map After being compressed into a single channel by 1×1 convolution, the initial prediction map is obtained by upsampling and applying a sigmoid activation function. Subsequently, feature maps are introduced. Global features and input image Enhance high-frequency target signals and output prediction maps :

[0030] in, For the first The enhanced output prediction map at each scale For the first Initial prediction maps at various scales, For the first Decoding feature maps at various scales, For the input image, This indicates dynamic high-frequency enhancement operation.

[0031] Through steps S100 to S300, this embodiment constructs a complete two-stage dynamic frequency domain modulation framework. In the first stage, spatial-frequency fusion and adaptive bandgap enhancement at the feature extraction end address the problems of insufficient global perception and limited feature representation. In the second stage, prediction guidance and dynamic high-frequency correction at the decoding and prediction end address the problems of low saliency of weak targets and blurred boundaries. The two stages work together to comprehensively improve the performance of image target segmentation from feature representation to result prediction.

[0032] The specific implementation method of adaptive band enhancement during feature extraction is further explained in detail. The process of dynamically generating a band filter kernel and using it to adaptively enhance the fused features to obtain enhanced features includes: performing global pooling on the fused features to obtain global context information; predicting band modulation parameters based on the global context information, including center frequency, bandwidth, and shape parameters; constructing a band filter kernel based on the band modulation parameters; transforming the fused features to the frequency domain, performing frequency domain modulation using the band filter kernel, inversely transforming it back to the spatial domain, and then superimposing and fusing it with the fused features to obtain the enhanced features.

[0033] like Figure 2 As shown, the fusion of spatial and frequency domains is achieved through a parallel dual-branch structure. The spatial branch extracts local spatial features through convolution operations, while the frequency branch captures global frequency domain features through discrete Fourier transform. The two outputs are then concatenated to form a fused feature that combines local precision with global generalization. Its specific mathematical expression is as follows: Spatial features are extracted using the following formula. :

[0034] Extract frequency domain features using the following formula. :

[0035] The spatial domain branch primarily extracts detailed information such as edges and textures, while the frequency domain branch mainly extracts global contextual information. To leverage the complementary nature of these two branches and obtain more discriminative feature representations, this invention concatenates the outputs of the spatial and frequency domain branches along the channel dimension and fuses them using a 3×3 convolution and ReLU activation function. The following formula fuses the spatial and frequency domain features to obtain the fused feature. :

[0036] in, and These are the output feature maps for the spatial domain branch and the frequency domain branch, respectively. To output the fused feature map, H and W represent the height and width of the input feature map, respectively. This represents the number of channels in the input feature map. This indicates the number of channels in the output feature map. For the input feature map, This represents a 3×3 convolution operation. To modify the activation function of the linear unit, For channel-by-channel Discrete Fourier Transform This is a channel-by-channel inverse discrete Fourier transform. It is a learnable frequency-domain complex filter kernel. This indicates element-wise multiplication. This indicates a channel splicing operation.

[0037] In this embodiment, global pooling is preferably performed using a combination of global average pooling and global max pooling. Concatenating the results of these two pooling methods along the channel dimension constructs a global contextual feature representation that combines a global overview with salient details. The mathematical expression of the above global pooling process is as follows:

[0038] in, For the input feature map, This indicates a global average pooling operation. This represents the global maximum pooling operation. This is the concatenated global description vector.

[0039] Then, the prediction of the first step is achieved through a fully connected layer and an activation function. The bandwidth enhancement parameters of each channel, i.e. , and And construct the first according to the following formula Gaussian bandgap enhancement filter core for each channel , The filter kernel elements are represented as:

[0040] In this formula, each parameter has a clear physical meaning and adjustment mechanism. Indicates the frequency magnitude. Center frequency parameter. The center frequency of the filter passband is determined when When the density is large, the filter kernel focuses on high-frequency detail regions, which is beneficial for enhancing edges and small targets; when... When the bandwidth is small, it focuses on the low-frequency region, which helps smooth noise or extract global background structure. (Bandwidth parameter) It controls the coverage area of ​​the passband, and has a smaller... This means that narrowband filtering has strong frequency selectivity and is suitable for separating features at specific scales; larger... This means broadband filtering, which can preserve a wider range of frequency components. Shape parameters The transition characteristics of the passband edge were adjusted when When it degenerates into a standard Gaussian function, the transition is smooth; when At this time, the passband edge becomes steeper, approaching an ideal rectangular window, enabling more precise truncation of ineffective frequency bands; when At that time, the edges became smoother. Through , and By dynamically adjusting these three parameters, the network can adaptively extract and enhance the frequency bands containing both target and background features based on the input features.

[0041] After constructing the frequency band filter kernels for each channel, the frequency band filter kernels of each channel are concatenated along the channel dimension to obtain the final frequency band filter kernel matrix, the mathematical expression of which is as follows:

[0042] in, This is the final bandpass filter kernel matrix after splicing. For the first Bandwidth filtering core for each channel, The number of channels in the output feature map is denoted by . Therefore, the generated bandpass filter kernel is content-adaptive, meaning that for each channel of the input feature map, the network can dynamically adjust the filtering strategy according to its unique spectral characteristics, thus overcoming the limitation of traditional methods that use fixed-parameter filters, which cannot adapt to complex and variable scenarios.

[0043] Subsequently, the fused features are transformed to the frequency domain using a Discrete Fourier Transform (DFT), and then multiplied element-wise with the frequency band filter kernel matrix to complete frequency domain modulation. Finally, an Inverse Discrete Fourier Transform (IFT) is used to restore the features to the spatial domain. To maintain training stability and preserve the fundamental information in the original features, this embodiment superimposes and fuses the frequency-enhanced features with the original fused features (i.e., residual connection) to obtain the enhanced features. The mathematical expression of this frequency band enhancement process is as follows:

[0044] in, This is the feature map after frequency band enhancement. As a feature of fusion, For channel-by-channel Discrete Fourier Transform This is a channel-by-channel inverse discrete Fourier transform. This is the frequency band filter kernel matrix. This indicates element-wise multiplication.

[0045] Furthermore, after obtaining the enhanced features, the method further includes processing the enhanced features through a channel attention module and a spatial attention module to suppress background features and enhance the target response, resulting in the final output features. Specifically, the channel attention module analyzes the feature correlation between channels and adaptively recalibrates the weights of each channel to enhance feature representation capabilities; the spatial attention module further focuses on the feature signals of the target and background by analyzing long-distance feature dependencies in the spatial dimension. The two modules are executed in series, significantly improving the signal-to-noise ratio and discriminative power of the feature representation. It should be understood that although this embodiment describes the execution order of channel-first, then spatial, other embodiments may also adopt the order of spatial-first, then channel, or a parallel execution followed by fusion approach, as long as multi-dimensional feature refinement can be achieved, all of which fall within the scope of protection of this invention. After introducing the attention mechanism, the mathematical expression of the overall feature transformation process is as follows:

[0046] in, The final output feature map, This indicates the channel attention module. This indicates a spatial attention module. This is the feature map after frequency band enhancement.

[0047] The dynamic high-frequency enhancement and correction mechanism in the decoding and prediction stage is further explained in detail. A high-pass filter kernel is dynamically generated, and the input image is enhanced using the initial prediction image to obtain a guide image. The guide image is then subjected to high-frequency filtering using the high-pass filter kernel, and the initial prediction image is corrected based on the high-frequency filtering result to obtain the target segmentation result.

[0048] like Figure 3 As shown, the decoded features contain semantic information after multiple layers of abstraction, reflecting the distribution of the target and background at the current scale. To accurately match this state, this embodiment first performs global pooling on the decoded features to obtain global context information. For example, global average pooling and global maximum pooling can be used simultaneously, and the two can be concatenated to form a global description vector. Subsequently, the global description vector is mapped to high-pass modulation parameters, including bandwidth and shape parameters, through a fully connected layer and the Softplus activation function. This process establishes a direct mapping from high-level semantics to frequency domain control parameters, enabling the generated high-pass filter kernel to adaptively adjust according to the current feature distribution. When the decoded features indicate a weak target or strong background clutter, the network can automatically adjust the parameters to enhance the extraction capability of specific high-frequency components; conversely, it maintains a relatively smooth transition to avoid introducing too much noise. The mathematical expression for generating the global description vector is as follows:

[0049] in, To decode the feature map, This indicates a global average pooling operation. This represents the global maximum pooling operation. This is the concatenated global description vector, used to predict high-pass modulation parameters through a fully connected layer and the Softplus activation function.

[0050] Furthermore, the specific mathematical expression for constructing a high-pass filter kernel based on high-pass modulation parameters is as follows: The high-pass filter kernel is constructed according to the following formula. :

[0051] in, For frequency domain coordinates, The distance from the frequency domain coordinates to the center point in the frequency domain. This is a bandwidth parameter used to control the high-pass filter core. The bandwidth; The shape parameter is used to control the high-pass filter kernel. The steepness of the decay.

[0052] In this formula, the high-pass filter kernel is constructed as a "1 minus low-pass form" structure. The exponential term is essentially a generalized Gaussian low-pass filter, used to characterize the degree of low-frequency component retention; subtracting this low-pass component from 1 mathematically guarantees that the gain at the frequency center (zero frequency) is 0, while the gain gradually approaches 1 as the frequency increases, thus achieving the high-pass filtering characteristics. Bandwidth parameter The location of the cutoff frequency is determined. The smaller the size, the wider the passband, and more mid-to-low frequency components are retained, which is beneficial for maintaining the outline and structure of larger targets; The larger the value, the narrower the passband, retaining only extremely high-frequency edge and noise information, making it suitable for extracting the fine contours of tiny targets. Shape parameters This controls the transition slope at the passband edge, resulting in a larger slope. A value that makes the filter kernel approach an ideal high-pass filter, exhibiting sharp cutoff characteristics, is beneficial for separating adjacent spectral components; a smaller value... The value provides a smooth transition, helping to reduce spatial ringing effects caused by frequency domain truncation. Through the dynamic combination of these two parameters, the network can customize the optimal high-frequency extraction strategy for each input image and even each decoding level.

[0053] In this embodiment, a key preliminary step is to enhance the input image using the initial prediction image to obtain a guidance image. Specifically, this step is achieved by performing element-wise operations between the initial prediction image and the input image. A preferred calculation method is to multiply the initial prediction image and the input image element-wise, and then add the result back to the original input image. ,in This represents the initial prediction map. Indicates the input image. This indicates element-wise multiplication. The reason for designing this specific guided enhancement step, rather than directly high-pass filtering the original input image, is that the original input image typically contains a large amount of background clutter and sensor noise. Direct filtering easily misinterprets these non-target high-frequency interferences as target signals, leading to an increased false alarm rate. Furthermore, it is not advisable to filter only the initial prediction image itself, because the initial prediction image has undergone multiple downsampling and nonlinear transformations. While it possesses good semantic localization capabilities, it often loses original texture details and weak high-frequency responses. The operation essentially uses the initial prediction map, expressed in probabilistic form, as a spatial attention mask to perform weighted filtering on the original input image. It preserves only the original high-frequency details in high-probability regions predicted as targets, while significantly suppressing the original signal in background regions. The subsequent... The operation introduces residual connections to ensure that even in regions with low prediction confidence, the original basis information is not completely lost, leaving room for subsequent corrections. This "prediction-guided + original detail injection" strategy effectively resolves the contradiction between "low signal-to-noise ratio" and "detail loss" in weak target segmentation, providing a high-purity target enhancement signal source for subsequent high-pass filtering.

[0054] After obtaining the guiding image, it is converted to the frequency domain, and then high-frequency filtering is performed using the aforementioned dynamically generated high-pass filter kernel. Finally, it is inversely transformed back to the spatial domain to obtain high-frequency enhanced features.

[0055]

[0056] in, This is the feature map after high-frequency enhancement. It is a high-pass filter core. To guide the enhanced image, For channel-by-channel Discrete Fourier Transform This is a channel-by-channel inverse discrete Fourier transform. This indicates element-wise multiplication. This high-frequency enhancement feature centrally reflects the target edge and texture information after denoising and enhancement.

[0057] Finally, the initial prediction map is weighted and corrected based on high-frequency enhanced features and the Sigmoid activation function.

[0058] in, This is the output prediction diagram after dynamic high-frequency enhancement. This is the initial prediction map. This is a high-frequency enhanced feature map, where Sigmoid is the Sigmoid activation function. This indicates element-wise multiplication.

[0059] As a preferred implementation, the correction process specifically includes: inputting high-frequency enhanced features into the Sigmoid activation function to obtain a Sigmoid weight map; multiplying the initial prediction map element-wise with the Sigmoid weight map, and superimposing the multiplication result with the initial prediction map to obtain the target segmentation result.

[0060] In this correction mechanism, the Sigmoid activation function plays a crucial role. It non-linearly maps the numerical range of high-frequency enhancement features to between 0 and 1, generating a normalized Sigmoid weight map. This means that high-frequency enhancement features are no longer directly added as additive terms to the prediction map, but are transformed into a "confidence adjustment factor." When the high-frequency response of a region is strong and matches the target feature, the corresponding Sigmoid weight approaches 1, and the value of the initial prediction map in that region is significantly enhanced; when the high-frequency response of a region is weak or exhibits noise characteristics, the Sigmoid weight approaches 0, and the value of the initial prediction map in that region remains basically unchanged or is slightly suppressed. The final superposition operation... This constitutes an adaptive feedback correction loop. Compared to simple linear superposition, this Sigmoid correction mechanism has better robustness. It can dynamically adjust the correction amplitude according to the reliability of high-frequency enhanced features, effectively avoiding artifacts introduced by over-enhancement or missed detections caused by under-enhancement. Thus, while ensuring a low false alarm rate, it significantly improves the detection capability of weak targets and the fineness of segmentation boundaries.

[0061] The overall network architecture and training strategy of the image target segmentation method are further explained in detail. The process of processing the input image through a feature extraction model to obtain multi-level feature maps specifically includes: at the encoding end, multiple downsampling and feature extraction operations are performed to extract multi-level semantic features of the input image layer by layer. Specifically, the encoding end typically consists of multiple cascaded feature extraction blocks, each containing a convolutional layer, an activation function layer, and a downsampling layer. As the number of network layers increases, the spatial resolution of the feature maps gradually decreases, while the semantic abstraction level gradually increases. This hierarchical feature extraction method can simultaneously capture shallow edge texture details and deep category semantic information, providing a rich multi-scale representation foundation for subsequent target segmentation. It should be understood that although this embodiment uses progressive downsampling as an example, in other embodiments, dilated convolutions or Transformers, which do not change the spatial resolution, can also be used to expand the receptive field. Any method that can achieve multi-level feature extraction falls within the scope of this invention.

[0062] Accordingly, the process of decoding based on multi-level feature maps to obtain decoded features and initial prediction maps specifically includes: at the decoding end, restoring the feature map resolution through multiple upsampling operations, and concatenating and fusing the corresponding level of encoding end features with the decoding end features to generate multi-scale features; generating initial prediction maps corresponding to each scale based on the multi-scale features. Specifically, the decoding end gradually restores spatial details through an upsampling path symmetrical to the encoding end. The key lies in the skip connection mechanism, which concatenates the encoding end feature maps and decoding end feature maps of the same scale in the channel dimension. This channel concatenation operation can complement and fuse the fine spatial structure preserved by the encoding end with the high-level semantic information restored by the decoding end, effectively alleviating the problem of target position information loss caused by multiple downsampling. In this embodiment, the concatenated features will also undergo the spatial-frequency domain fusion enhancement processing described in the previous embodiment, so that the cross-level transmitted features not only have spatial alignment, but also undergo adaptive feature enhancement in the frequency domain dimension. Subsequently, based on the fused multi-scale features, initial prediction maps corresponding to each scale can be generated through independent prediction branches. These prediction maps at different scales correspond to target responses of different sizes, providing multi-granularity candidate criteria for the final accurate segmentation.

[0063] Furthermore, to integrate multi-scale prediction information to obtain a unified segmentation result, the process of correcting the initial prediction map based on the high-frequency filtering result to obtain the target segmentation result also includes: channel stitching of the corrected initial prediction maps corresponding to each scale, and processing through convolution operations and activation functions to output the final binary classification segmentation result map. Specifically, after the dynamic high-frequency enhancement correction described in the above embodiments, the prediction maps at each scale already contain the enhanced target high-frequency signal. These corrected prediction maps are stitched together again in the channel dimension, and 1×1 convolution is used for cross-channel information interaction and dimensionality reduction. Finally, activation functions such as Sigmoid are used to map pixel values ​​to the probability range of 0 to 1, thereby obtaining a binary classification segmentation result map with the same size as the input image. This "scale-wise correction-unified fusion output" strategy ensures both the saliency enhancement of small targets at specific scales and the continuity and integrity of the final result in the global space.

[0064] Furthermore, to guide the parameter learning of the aforementioned network model, this invention employs the Intersection over Union (IoU) loss function as the optimization objective during the training phase. The specific formula for calculating the IoU loss function is as follows:

[0065] in, Indicates the total number of pixels. Indicates the first The predicted probability value of each pixel. Indicates the first The actual label value of each pixel. For smoothing. Introduce a smoothing term. (For example, the value is 10) -7 This method offers two key advantages: firstly, it prevents numerical instability caused by zero denominators; secondly, in the early stages of training when the predicted results do not overlap with the true labels, the smoothing term ensures the continuity of the loss function's gradient, avoiding the vanishing gradient problem and enabling the model to smoothly begin the learning process. Compared to traditional cross-entropy loss, IoU loss directly optimizes the core evaluation metric of the segmentation task and exhibits better convergence performance in weak target segmentation scenarios with extremely imbalanced positive and negative samples.

[0066] Regarding model training configuration, this embodiment preferably uses the Adam optimizer for parameter updates. The Adam optimizer combines momentum and adaptive learning rate adjustment mechanisms, enabling it to dynamically adjust the learning rate of each parameter based on the first and second moment estimates of the gradient. In specific implementation, the initial learning rate can be set to 10. -4 It also employs cosine annealing or step decay strategies for learning rate scheduling, with the weight decay coefficient set to 10. -5 This is to prevent overfitting. This optimized configuration can accelerate the convergence speed of the model on complex non-convex loss surfaces while ensuring training stability, and ensure that various learnable parameters (such as band modulation parameters, high-pass modulation parameters, etc.) in the aforementioned dynamic frequency domain modulation mechanism can be fully and effectively optimized.

[0067] like Figures 1 to 3 As shown, this embodiment provides an infrared small target segmentation device, which is a concrete manifestation of the aforementioned method embodiment in terms of product form. It implements a two-stage dynamic frequency domain modulation mechanism through functional modularization. The device includes: an image acquisition module, a hybrid filtering module, and a dynamic high-frequency enhancement module. It should be understood that these modules can be software functional units running on a processor, hardware logic circuits integrated into a dedicated chip, or a hybrid implementation combining software and hardware, as long as they can execute the following data processing logic.

[0068] The image acquisition module is used to acquire input images. Specifically, this module serves as the data entry point for the entire segmentation device, responsible for receiving raw image data from external acquisition devices or storage media. In addition to basic data reading functions, this module typically includes a preprocessing unit to perform operations such as normalization, size standardization, or denoising on the input image, ensuring that the data received by subsequent modules conforms to the input specifications of the feature extraction model. The standardized input image output by this module is simultaneously transmitted to both the hybrid filtering module and the dynamic high-frequency enhancement module. The former is used for feature encoding, and the latter for subsequent guided enhancement processing, thereby establishing a direct pathway between the raw data and high-level semantic processing.

[0069] The hybrid filtering module processes the input image using a feature extraction model to obtain multi-level feature maps. During feature extraction, spatial and frequency domain fusion processing is performed on the extracted features to obtain fused features, and a frequency band filter kernel is dynamically generated. This kernel is then used to adaptively enhance the fused features, resulting in enhanced features. Specifically, this module internally corresponds to the encoding stage in the method embodiment. It contains cascaded feature extraction subunits and adaptive frequency band modulation subunits. The feature extraction subunits construct a multi-level feature pyramid through multiple downsampling and convolution operations; the adaptive frequency band modulation subunit is embedded in the feature extraction path, receiving the fused features of the current level in real time. Within this subunit, the data stream is first compressed into a global vector through a global pooling layer, which then drives the parameter prediction network to generate the frequency band filter kernel. Subsequently, the fused features are transformed to the frequency domain and interact with the filter kernel, then inversely transformed back to the spatial domain to complete the enhancement. This modular encapsulation ensures that the deep fusion of spatial local perception and frequency global perception can be automatically executed at each level of feature extraction without external intervention, thus guaranteeing the robustness of feature representation.

[0070] The dynamic high-frequency enhancement module is used to decode features and an initial prediction map based on multi-level feature maps. It dynamically generates a high-pass filter kernel, uses the initial prediction map to enhance the input image to obtain a guide image, applies high-frequency filtering to the guide image using the high-pass filter kernel, and corrects the initial prediction map based on the high-frequency filtering result to obtain the target segmentation result. Specifically, this module corresponds to the decoding and optimization stage in the method embodiment. It integrates an upsampling recovery unit, a guide enhancement unit, and a high-frequency feedback correction unit. The upsampling recovery unit is responsible for gradually restoring the spatial resolution and generating the initial prediction map; the guide enhancement unit receives the original input image and the current initial prediction map from the image acquisition module, and generates a guide image with a higher signal-to-noise ratio through element-wise operations; the high-frequency feedback correction unit processes the guide image using a high-pass filter kernel dynamically generated based on the decoding features, and feeds back the extracted high-frequency enhancement signal to the initial prediction map through a gating mechanism. This module forms a tight closed-loop data flow, namely "prediction-guide-enhancement-correction," enabling targeted improvement of the segmentation accuracy of weak targets at the decoding end. Through the coordinated work of the above modules, the device fully reproduces the core concept of the method of the present invention, and effectively solves the technical problems of limited target feature extraction and insufficient prediction significance in complex backgrounds while ensuring processing efficiency.

[0071] This embodiment provides a computer device and a computer-readable storage medium for providing a hardware execution environment and software distribution carrier for the aforementioned image target segmentation method. It should be understood that the computer device and storage medium described in this embodiment are concrete manifestations of the aforementioned method embodiments at the physical entity and data product level. Their core improvement lies in carrying and executing image processing logic including a two-stage dynamic frequency domain modulation mechanism.

[0072] As one implementation method, this embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the image target segmentation method described in any of the foregoing method embodiments.

[0073] Specifically, the processor can be a central processing unit (CPU), or a computing chip with data processing capabilities such as a graphics processing unit (GPU), neural network processing unit (NPU), field-programmable gate array (FPGA), or application-specific integrated circuit (ASIC). Considering that this invention involves a large number of frequency domain transformations (such as discrete Fourier transform and its inverse transform), matrix convolution operations, and tensor operations in attention mechanisms, a GPU or NPU is preferably used as the main processor or coprocessor. Its massively parallel computing architecture accelerates the feature extraction and frequency band enhancement processes to meet the performance requirements of real-time segmentation. The memory can include high-speed random access memory (RAM) for caching input images, multi-level feature maps, frequency band filtering kernel parameters, and intermediate activation values ​​during inference; it can also include non-volatile memory (such as flash memory or solid-state drives) for persistently storing trained feature extraction model weights, parameters of the frequency band modulation parameter prediction network, and operating system and application code. The processor and memory communicate via a system bus, which can be an address bus, data bus, or control bus, responsible for efficiently transmitting massive amounts of image data and control commands between components. When the processor reads and executes the computer program from the memory, it instantiates a series of processing flows described in the above embodiments, such as image acquisition, space-frequency fusion feature extraction, adaptive frequency band enhancement, decoding prediction guidance, and dynamic high-frequency correction, at the hardware level. This transforms the abstract algorithm logic into specific physical signal processing actions, enabling accurate segmentation of weak targets in complex backgrounds.

[0074] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the infrared small target segmentation method described in any of the foregoing method embodiments.

[0075] Specifically, a computer-readable storage medium can be any tangible device capable of containing, storing, transmitting, or propagating program code for use by an instruction execution system. For example, the medium can include, but is not limited to, devices with electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, or any combination of the above systems. More specific examples include portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disc read-only memory (CD-ROM), digital versatile optical disc (DVD), or other optical storage devices. In this invention, the computer program stored on the storage medium contains a series of machine-readable instructions that, when loaded into the processor of a computer device and executed, precisely control hardware resources to perform operations according to the timing and logic defined in the above embodiments. For example, the instruction sequence first configures an image interface to acquire an input image, then calls mathematical library functions or a dedicated hardware accelerator to perform spatial and frequency domain fusion processing, dynamically constructs filter kernel parameters based on global feature information generated during runtime, and finally outputs the target segmentation result. By embedding the technical solution of this invention into a storage medium, it not only facilitates the deployment, distribution, and version management of the algorithm model, but also provides a clear product subject basis for the exercise of patent rights. It should be understood that regardless of the specific physical form of the storage medium, as long as the program instructions stored therein can implement the infrared small target segmentation method described in this invention when executed, they fall within the protection scope of this embodiment.

[0076] like Figure 1 As shown, this embodiment provides a specific application scenario example of applying the aforementioned infrared small target segmentation method to small target segmentation in infrared remote sensing images. It should be understood that this embodiment is only used to demonstrate the effectiveness and practicality of the technical solution of the present invention in actual engineering problems and does not constitute a limitation on the scope of protection of the aforementioned embodiments. In other embodiments, the present invention is also applicable to image processing tasks with similar "weak target, strong background" characteristics, such as lesion detection in medical imaging and micro-defect recognition in industrial vision.

[0077] In infrared remote sensing scenarios, the input image is typically a single-channel infrared grayscale image. Its typical characteristics include an extremely small target imaging area (often occupying only a few to tens of pixels), a lack of structured information such as shape and texture, and easy obstruction by complex background clutter such as clouds and surface thermal radiation, resulting in an extremely low signal-to-noise ratio. For this specific scenario, this embodiment provides a concrete configuration and verification of the aforementioned general image target segmentation method.

[0078] Specifically, in this application scenario, the input image is set as a single-channel infrared remote sensing image with a resolution of 256×256. The feature extraction model adopts a four-layer encoder-decoder symmetrical structure to adapt to the multi-scale feature extraction requirements at this resolution. At the encoding end, each layer embeds the hybrid filtering module described in the previous embodiment. When the infrared image is input, the spatial branch extracts local spatial features such as edge contours through 3×3 convolution; at the same time, the frequency domain branch transforms the image to the frequency domain through discrete Fourier transform to capture global distribution features. Addressing the differences in frequency distribution between infrared small targets and background clutter, the adaptive band enhancement mechanism plays a crucial role: based on the global context information of the fused features at the current level, the system dynamically predicts the center frequency parameters, bandwidth parameters, and shape parameters that can effectively enhance the target feature band and background feature band, constructing a personalized band filtering kernel. For example, when processing infrared images containing dense clouds, the network can automatically adjust the center frequency and bandwidth of the frequency band, accurately enhance the high-frequency components corresponding to small target scales, and suppress the low-frequency background energy corresponding to large cloud areas and the ground surface, thereby improving the separability of targets and backgrounds at the source of feature extraction.

[0079] In the decoding and prediction phase, the dynamic high-frequency enhancement module is specifically optimized to address the false alarm problem in infrared scenes. Due to the complex background of infrared images, conventional decoding and prediction are prone to generating numerous false responses caused by cloud edges or noise. In this embodiment, the initial prediction map is used to perform guided enhancement (i.e., execute...) on the original infrared input image. The first step (calculation) has special physical significance in infrared scenes: it's equivalent to using a preliminary semantic-level judgment as a mask to "extract" the true signal of the potential target area from the raw thermal radiation data, while significantly suppressing the original noise in non-target areas. Subsequently, a high-pass filter kernel dynamically generated based on the decoded features processes the guiding image. The bandwidth and shape parameters of this filter kernel change in real time according to the semantic content of the decoded features. For example, when the decoded features identify the current area as a smooth sea surface background, the high-pass filter kernel automatically adjusts to a steeper cutoff characteristic to completely filter out low-frequency undulations in the sea surface; while when the current area is identified as a complex land background, it adjusts to a relatively gentle transition characteristic to avoid interfering with weak target signals overlapping with the background spectrum. Finally, the high-frequency enhancement result is fed back to correct the initial prediction map through the Sigmoid activation function, forming a closed-loop optimization. This mechanism effectively distinguishes the high-frequency thermal radiation signal of a real small target from the pseudo-high-frequency interference of background clutter, significantly reducing the false alarm rate.

[0080] Regarding model training configuration, this embodiment uses the Intersection over Union (IoU) loss function as the optimization objective and sets a smoothing term. 10 -7To ensure numerical stability, the optimizer Adam was chosen, with an initial learning rate set to 10. -4 The weight decay factor is set to 10. -5 The batch size was set to 16. After training and testing on a publicly available infrared small target dataset, compared to traditional segmentation networks that rely solely on spatial convolution, this scheme maintains a high recall rate while effectively reducing the false alarm rate. Especially under conditions of strong clutter edges and extremely low signal-to-noise ratios, it significantly improves the detection capability and boundary localization accuracy for small targets. This fully demonstrates the superiority of the proposed dual-stage dynamic frequency domain modulation mechanism in solving the specific technical challenge of infrared small target segmentation.

[0081] It must be emphasized again that the above-described processing flow, number of network layers, parameter settings, and experimental data for infrared remote sensing images are merely specific examples to illustrate the feasibility of the technical solution of this invention. In practical applications, the resolution of the input image, the number of channels, the number of encoder layers, the specific frequency domain transformation method, and the training hyperparameters can all be adjusted according to specific task requirements, and these adjustments should not be considered as changes that depart from the scope of protection of this invention.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for segmenting small infrared targets, characterized in that, Includes the following steps: Obtain the input image; The input image is processed by a feature extraction model to obtain a multi-level feature map. During the feature extraction process, the extracted features are fused by spatial and frequency domains to obtain fused features. A frequency band filter kernel is dynamically generated, and the fused features are adaptively enhanced using the frequency band filter kernel to obtain enhanced features. Decoding is performed based on the multi-level feature map to obtain decoded features and an initial prediction map; A high-pass filter kernel is dynamically generated. The input image is then enhanced using the initial prediction image to obtain a guide image. The guide image is then subjected to high-frequency filtering using the high-pass filter kernel. Based on the high-frequency filtering result, the initial prediction image is corrected to obtain the target segmentation result.

2. The method according to claim 1, characterized in that, The dynamically generated frequency band filter kernel is used to enhance the frequency band of the fused features to obtain enhanced features, including: The fused features are subjected to global pooling to obtain global context information; Based on the global context information, frequency band modulation parameters are predicted, including center frequency parameters, bandwidth parameters, and shape parameters. The frequency band filter kernel is constructed based on the frequency band modulation parameters; The fusion feature is transformed to the frequency domain, and after frequency domain modulation using the frequency band filter kernel, it is inversely transformed back to the spatial domain and superimposed and fused with the fusion feature to obtain the enhanced feature.

3. The method according to claim 2, characterized in that, The construction of the frequency band filter kernel based on the frequency band modulation parameters includes: Construct the first according to the following formula Each channel's band filter core : in, For frequency domain coordinates, The Euclidean distance from the frequency domain coordinates to the center point of the frequency domain is used to characterize the frequency magnitude; The center frequency parameter, The bandwidth parameters are used to control the bandwidth. The shape parameters are used to control the steepness of the attenuation. The frequency band filter kernels of each channel are concatenated along the channel dimension to obtain the final frequency band filter kernel matrix.

4. The method according to claim 1, characterized in that, After obtaining the enhanced features, the method further includes: The enhanced features are processed by the channel attention module and the spatial attention module to suppress background features and enhance the target response, resulting in the final output features.

5. The method according to claim 1, characterized in that, The process involves dynamically generating a high-pass filter kernel, using the initial prediction image to enhance the input image to obtain a guide image, applying the high-pass filter kernel to the guide image for high-frequency filtering, and correcting the initial prediction image based on the high-frequency filtering result to obtain the target segmentation result, including: The decoded features are subjected to global pooling to obtain global context information; High-pass modulation parameters are predicted based on the global context information, and the high-pass modulation parameters include bandwidth parameters and shape parameters; The high-pass filter kernel is constructed based on the high-pass modulation parameters; The initial prediction image and the input image are processed element-wise to obtain the guiding image; The guiding image is converted to the frequency domain, and after high-frequency filtering using the high-pass filter kernel, it is inversely transformed back to the spatial domain to obtain high-frequency enhanced features. The initial prediction map is weighted and corrected based on the high-frequency enhanced features and the Sigmoid activation function to obtain the target segmentation result.

6. The method according to claim 5, characterized in that, The construction of the high-pass filter kernel based on the high-pass modulation parameters includes: Construct a high-pass filter kernel according to the following formula. : in, For frequency domain coordinates, The distance from the frequency domain coordinates to the center point of the frequency domain. This is a bandwidth parameter used to control the high-pass filter core. The bandwidth; The shape parameter is used to control the high-pass filter kernel. The steepness of the decay.

7. The method according to claim 5, characterized in that, The step of weighting and correcting the initial prediction map based on the high-frequency enhanced features and the Sigmoid activation function to obtain the target segmentation result includes: The high-frequency enhancement features are input into the Sigmoid activation function to obtain the Sigmoid weight map; The initial prediction map is multiplied element-wise with the Sigmoid weight map, and the multiplication result is superimposed on the initial prediction map to obtain the target segmentation result.

8. The method according to claim 1, characterized in that, The process of processing the input image using a feature extraction model to obtain multi-level feature maps includes: At the encoding end, multiple downsampling and feature extraction operations are performed to extract multi-level semantic features of the input image layer by layer. The decoding process based on the multi-level feature maps to obtain decoded features and an initial prediction map includes: At the decoding end, the feature map resolution is restored through multiple upsampling operations, and the corresponding level of encoding end features and decoding end features are concatenated and fused to generate multi-scale features; The initial prediction map corresponding to each scale is generated based on the multi-scale features.

9. The method according to claim 1, characterized in that, The step of correcting the initial prediction map based on the high-frequency filtering results to obtain the target segmentation result includes: Channel stitching is performed on the corrected initial prediction maps corresponding to each scale, and the final binary classification segmentation result map is output through convolution operation and activation function processing.

10. An infrared small target segmentation device, characterized in that, include: The image acquisition module is used to acquire the input image; The hybrid filtering module is used to process the input image through a feature extraction model to obtain multi-level feature maps; During the feature extraction process, the extracted features are fused in both the spatial and frequency domains to obtain fused features. A frequency band filter kernel is dynamically generated, and the fused features are then adaptively enhanced using the frequency band filter kernel to obtain enhanced features. The dynamic high-frequency enhancement module is used to perform decoding processing based on the multi-level feature map to obtain decoded features and an initial prediction map; A high-pass filter kernel is dynamically generated. The input image is then enhanced using the initial prediction image to obtain a guide image. The guide image is then subjected to high-frequency filtering using the high-pass filter kernel. Based on the high-frequency filtering result, the initial prediction image is corrected to obtain the target segmentation result.

Images