Progressive camouflage target detection method based on dual-domain fusion

Abstract

The application discloses a kind of progressive camouflage target detection methods based on dual-domain fusion, obtains and divides public data set;Progressive frequency domain-spatial domain collaborative optimization network is constructed, including encoder, frequency band axial attention module, amplitude guide spatial modulator and multi-scale adaptive fusion module, frequency band axial attention module fuses discrete wavelet transform with axial attention to capture directional local structure clues, amplitude guide spatial modulator introduces learnable frequency domain attention mechanism to realize the adaptive modulation of spatial features in frequency domain, multi-scale adaptive fusion module is fused by pixel-level weighting multi-scale features and introduces edge auxiliary supervision;End-to-end optimization is carried out to network using training set and total loss function, and total loss function combines regional segmentation loss and edge auxiliary supervision loss;Finally, the performance of the network is evaluated using evaluation indicators, realizes the deep interaction of spatial and frequency domain information, improves the detection accuracy and boundary integrity of camouflage target.

Description

Technical Field

[0001] This invention belongs to the field of computer target vision detection technology, and in particular relates to a progressive camouflage target detection method based on dual-domain fusion. Background Technology

[0002] This paper focuses on COD (Camouflaged Object Detection), a cutting-edge topic in computer vision. Its goal is to identify and segment targets that blend seamlessly with the background environment in terms of color, texture, and pattern, making them difficult to detect with the naked eye. This technology has broad application value in fields such as medical image analysis, industrial defect detection, and ecological research.

[0003] Traditional detection methods primarily rely on hand-designed features, which have limited generalization ability when facing complex natural scenes. With the development of deep learning technology, existing COD methods can be mainly summarized into the following two technical approaches: (1) Detection methods based on spatial domain features. These methods mainly use CNN or Transformer-based backbone networks to extract and fuse multi-scale features from the spatial domain of the image. By enhancing context modeling, edge priors, or attention mechanisms, they improve the model's ability to perceive the overall structure and local contours of the target. (2) Methods that introduce frequency domain information as auxiliary cues. In order to overcome the limitations of spatial representation in capturing subtle textures and global structures, some studies have attempted to introduce frequency domain analysis, using the spectral features of the image as supplementary information, and splicing or interacting with spatial features to enhance the model's ability to discriminate the frequency domain differences between the target and the background.

[0004] However, existing methods still have limitations in complex scenes because camouflaged targets are highly similar to the background in terms of visual cues. Methods that rely solely on spatial domain features are difficult to use frequency domain discrimination cues, resulting in insensitivity to subtle structural flaws. Methods that introduce frequency domain information still face the following three key problems: (1) the coordination between spatial and frequency domain features is relatively shallow, failing to explicitly model and utilize directional structural information in the frequency band; (2) the decoding end lacks a continuous spectral domain calibration mechanism for target discrimination, failing to effectively map the spectral domain response to the spatial domain modulation signal; (3) the multi-scale feature fusion capability is insufficient, lacking cross-scale collaborative modeling and unified representation, and the edge prior is not fully integrated into the fusion process, resulting in weak structural consistency and easily blurred boundaries. Therefore, it is necessary to design a progressive camouflaged target detection method based on dual-domain fusion to solve the above problems. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide a progressive camouflage target detection method based on dual-domain fusion. This method addresses the issues of global structural instability and missing local details caused by the high homogeneity between the target and the background in camouflage target detection. By constructing a collaborative modeling and progressive optimization mechanism that integrates frequency and spatial domain information, the method enhances the ability to discriminate subtle differences between the camouflage target and the background. Furthermore, it progressively optimizes the consistency of the target structure and boundary details during the decoding process, thereby improving the detection and segmentation accuracy of camouflage targets in complex scenes.

[0006] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows: The progressive camouflage target detection method based on dual-domain fusion includes the following steps: S1, Data Preparation and Model Input: Obtain a public dataset, divide it into a training set and a test set, and preprocess the images in the training set; S2, Network Design and Construction: Construct an asymptotic frequency domain-spatial domain co-optimization network, which includes an encoder, a frequency band axial attention module, an amplitude-guided spatial modulator, and a multi-scale adaptive fusion module; The frequency band axial attention module integrates discrete wavelet transform and axial attention to capture directional local structural cues of camouflaged targets; Amplitude-guided spatial modulators introduce a learnable frequency-domain attention mechanism to adaptively reweight the amplitude spectrum and modulate frequency-domain information into spatial features; The multi-scale adaptive fusion module fuses multi-scale features through pixel-level adaptive weighted fusion and introduces an edge-assisted supervision branch; S3, Model Optimization: The preprocessed training set images are input into the progressive frequency-spatial co-optimization network. The network is trained and optimized end-to-end using a deep supervision strategy and a total loss function. The total loss function combines region segmentation loss and edge-assisted supervision loss. S4, Model Performance Evaluation: The performance of the trained network is evaluated using test set images, employing metrics including structural measures. Enhance alignment metrics Weighted metrics and The evaluation metrics, including those used in the study, are employed to verify the effectiveness of camouflage target detection.

[0007] Preferably, in step S1, the public dataset uses four public benchmark datasets: CAMO, COD10K, CHAMELEON, and NC4K; the training set consists of 1000 images from CAMO and 3040 images from COD10K, for a total of 4040 images; the test set consists of the complete CHAMELEON dataset, the complete NC4K dataset, and the test sets of CAMO and COD10K.

[0008] Furthermore, during training, the input images are uniformly scaled to a resolution of 512×512, and data augmentation is performed using random flipping, rotation, and cropping.

[0009] Preferably, in step S2, the band axial attention module (BAA) is configured to perform the following process: (1) For input features, the BAA module is at the deepest level Its input is only the feature map of the current level, while in other levels, it takes the BAA output feature map of the previous level. and the output characteristics of the current level SMT backbone network The inputs from two levels are fused using a gated fusion unit (GFU). The process is as follows: ; ; in, For element-wise multiplication, 3 3 convolutions, For batch normalization, express Activation function; (2) The BAA module performs two parallel processes simultaneously: Spatial axial features were extracted along the vertical and horizontal directions using adaptive average pooling AAP. and : ; Perform a two-dimensional discrete wavelet transform on the input features to extract frequency domain subbands; considering Subbands primarily capture the vertical edge and structural variations of an image. The sub-band primarily reflects edge and texture variations in the horizontal direction; it is projected using the axial frequency domain projection unit (AFP) to... and Subbands are processed to obtain frequency domain cues representing the vertical and horizontal directions, respectively. and ; (3) The spatial axial features and their corresponding frequency domain cues are spliced ​​together in the channel dimension and fused through the feature fusion unit CBR to generate fused features. and This models the correlation between the spatial dimension and the frequency domain dimension in a specific direction, and its expression is: ; ; in, CBR is the feature fusion unit, and the mapping unit is the mapping unit. This indicates channel-dimensional splicing, with AFP being the axial frequency domain projection unit. (4) Finally, attention weights in the vertical and horizontal directions are generated. and Weight and After channel-weighting the input features, they are further refined through a spatial attention module, and finally output through residual connections. : .

[0010] Through the aforementioned mechanism, the BAA module provides the network with crucial, direction-aware local geometric cues, improving the initial localization capability of camouflaged targets.

[0011] Preferably, in step S2, the amplitude-guided spatial modulator AGSM is configured to perform the following process: make Indicates the first The features after the hierarchy is processed by the BAA module, among which These correspond to four levels from deep to shallow; through feature projection and upsampling operations, the transfer of high-level semantic information to low-level features is achieved. ; in, This is a bilinear upsampling operation. This represents the multi-scale features after fusion; For the deepest layer ( ), based solely on cross-layer fusion features As input, output features For other levels ( The input of the AGSM module is determined by the output characteristics of the previous level AGSM module. Cross-layer fusion features with the current layer Assembled by splicing, and through The number of channels is adjusted by convolution, and it is denoted as... : ; First, the Fast Fourier Transform (FFT) is used to... Transform to the frequency domain to obtain The amplitude and phase spectra are obtained. After taking the logarithm of the amplitude spectrum to enhance stability, a lightweight frequency domain attention module (FA) is used to analyze the amplitude spectrum. The FA learns the importance of different frequency components through global average pooling and fully connected layers to generate adaptive modulation weights. The modulated amplitude spectrum is combined with the original phase spectrum and reconstructed by inverse Fourier transform (IFFT) to obtain the frequency domain enhanced features. The process is represented as follows: ; in, Represents grouped convolution. and These represent FFT and IFFT respectively, and FA represents the frequency domain attention module. Represents a unit complex number consisting of the original phase spectrum; Meanwhile, input features Spatial domain features are also extracted through a lightweight convolutional path. A gated fusion unit (GFU) is introduced to adaptively fuse frequency domain enhancement information and spatial structure information, generating channel-dimensional fusion weights based on the input. The final fusion features are: ; The fused features are output through residual connections and upsampling, serving as the refined features of the current level and passed to the next level.

[0012] Preferably, in step S2, the multi-scale adaptive fusion module MSAF is configured to perform the following process: The module input consists of feature maps at four scales. First, each feature map is projected to a uniform number of channels using a 1×1 convolution and downsampled to the same spatial size. Then, a depthwise separable convolutional enhancement block (EDSB) combined with channel attention (SEBlock) is used to enhance the features at each scale, thereby strengthening their discriminative power and suppressing redundancy, resulting in enhanced features. ; Four enhanced features spliced ​​along the channel dimension This is used to achieve more refined multi-scale feature fusion; Subsequently, the weight generation network WGNet predicts a pixel-level fusion weight for each scale feature at each spatial location. Weighting and learnable temperature coefficient Normalization to achieve finer fusion; the ultimate multi-scale unified representation. It is obtained by weighted summation of the enhancement features at each scale according to their respective weights: ; ; Characterization It also serves for subsequent prediction and monitoring: The features are upsampled and concatenated with the original features from each decoding stage, and then processed by the edge enhancement module. Processing is used to generate detailed prediction maps for each stage. ;at the same time An edge prediction map is generated using a lightweight edge prediction head, EH. And auxiliary supervision is provided by edge ground truth values ​​derived from the real mask: .

[0013] Preferably, step S3 includes end-to-end training and optimization using a deep supervision strategy: Prediction graphs of the four stages of the network decoder All are supervised by the region segmentation loss, which is composed of the binary cross-entropy loss. Compared with the intersection and comparison loss Together constitute; Introducing edge-assisted supervision to enhance boundary details, edge ground truth The edge prediction is obtained by processing the real mask with the Sobel operator and Gaussian filter. Generated by the MSAF module and using Dice loss. Supervision is performed; total loss function Defined as: ; in For the real mask, To balance the weights.

[0014] Preferably, in step S4, structural measurement The calculation method is as follows: ; in, This represents structural similarity based on target perception. This represents region-aware structural similarity. It is a hyperparameter, and its value range is... , The higher the value, the better the detection effect.

[0015] Preferably, in step S4, the alignment metric is enhanced. The calculation method is as follows: ; in, For the prediction map and the ground truth map at the pixel level Alignment balance at the location These are the height and width of the image, respectively; the calculated values ​​are... The higher the value, the better the detection effect.

[0016] Preferably, in step S4, the weighted metric... The calculation method is as follows: ; ; ; in, Indicates weighted precision. Indicates the weighted recall rate. To balance the importance of precision and recall; Represents pixels Spatial weights are used to measure the importance of a pixel in the target structure; , , These represent the true positive, false positive, and false negative cases at the pixel level; the calculated... The higher the value, the more accurate the test result.

[0017] Preferably, in step S4, the mean absolute error The calculation method is as follows: ; in, This represents the predicted probability or normalized pixel value. This represents the corresponding true pixel value. These are the height and width of the image, respectively; the calculated values ​​are... The smaller the value, the smaller the error and the better the detection effect.

[0018] The beneficial effects of this invention are as follows: Compared with existing methods, the Progressive Frequency-Spatial Co-optimization Network (PFSNet) proposed in this invention achieves deep interaction and optimization of spatial and frequency domain information at multiple levels by constructing a progressive collaborative mechanism of "directional cue capture—global spectral domain calibration—cross-scale discriminative refinement". This scheme can effectively discover subtle directional structural fractures, focus on the target discrimination region through adaptive frequency domain calibration, and utilize cross-scale fusion to stabilize structural consistency, thereby alleviating the contradiction between global and local features in camouflaged target detection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the model framework of the present invention; Figure 2 This is a schematic diagram of the frequency band axial attention module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the frequency band axial attention module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the scale-adaptive fusion module in an embodiment of the present invention; Figure 5 The structural metrics of each method in the embodiments of the present invention on four datasets. Schematic diagram of indicator bar chart. Detailed Implementation

[0020] Example 1: like Figure 1 As shown, the progressive camouflage target detection method based on dual-domain fusion includes the following steps: S1, Data Preparation and Model Input: Obtain a public dataset, divide it into a training set and a test set, and preprocess the images in the training set; S2, Network Design and Construction: Construct an asymptotic frequency domain-spatial domain co-optimization network, which includes an encoder, a frequency band axial attention module, an amplitude-guided spatial modulator, and a multi-scale adaptive fusion module; The frequency band axial attention module integrates discrete wavelet transform and axial attention to capture directional local structural cues of camouflaged targets; Amplitude-guided spatial modulators introduce a learnable frequency-domain attention mechanism to adaptively reweight the amplitude spectrum and modulate frequency-domain information into spatial features; The multi-scale adaptive fusion module fuses multi-scale features through pixel-level adaptive weighted fusion and introduces an edge-assisted supervision branch; S3, Model Optimization: The preprocessed training set images are input into the progressive frequency-spatial co-optimization network. The network is trained and optimized end-to-end using a deep supervision strategy and a total loss function. The total loss function combines region segmentation loss and edge-assisted supervision loss. S4, Model Performance Evaluation: The performance of the trained network is evaluated using test set images, employing metrics including structural measures. Enhance alignment metrics Weighted metrics and The evaluation metrics, including those used in the study, are employed to verify the effectiveness of camouflage target detection.

[0021] Preferably, in step S1, the public dataset uses four public benchmark datasets: CAMO, COD10K, CHAMELEON, and NC4K; the training set consists of 1000 images from CAMO and 3040 images from COD10K, for a total of 4040 images; the test set consists of the complete CHAMELEON dataset, the complete NC4K dataset, and the test sets of CAMO and COD10K.

[0022] Furthermore, during training, the input images are uniformly scaled to a resolution of 512×512, and data augmentation is performed using random flipping, rotation, and cropping.

[0023] Preferably, in step S2, during camouflage target detection, the high similarity in color and texture between the target and the background makes it difficult to capture initial localization clues based solely on spatial domain features. While existing methods incorporate frequency domain analysis, they often treat frequency domain information as overall auxiliary features, failing to explicitly mine and correlate the directional local structural breakage information contained within the frequency domain subbands. These subtle geometric flaws are crucial for distinguishing camouflage targets. Therefore, a frequency band axial attention module (BAA) is designed, with the structure as follows: Figure 2 As shown.

[0024] The core idea of ​​this module is to use Discrete Wavelet Transform (DWT) to analyze the frequency domain information of an image in different orthogonal directions, and combine it with spatial axial attention to enhance the sensitivity of features to structural fractures and texture anisotropy in specific directions.

[0025] The Bandwidth Axial Attention Module (BAA) is configured to perform the following process: (1) For input features, the BAA module is at the deepest level Its input is only the feature map of the current level, while in other levels, it takes the BAA output feature map of the previous level. and the output characteristics of the current level SMT backbone network The inputs from two levels are fused using a gated fusion unit (GFU). The process is as follows: ; ; in, For element-wise multiplication, 3 3 convolutions, For batch normalization, express Activation function; (2) The BAA module performs two parallel processes simultaneously: Spatial axial features were extracted along the vertical and horizontal directions using adaptive average pooling AAP. and : ; Perform a two-dimensional discrete wavelet transform on the input features to extract frequency domain subbands; considering Subbands primarily capture the vertical edge and structural variations of an image. The sub-band primarily reflects edge and texture variations in the horizontal direction; it is projected using the axial frequency domain projection unit (AFP) to... and Subbands are processed to obtain frequency domain cues representing the vertical and horizontal directions, respectively. and ; (3) The spatial axial features and their corresponding frequency domain cues are spliced ​​together in the channel dimension and fused through the feature fusion unit CBR to generate fused features. and This models the correlation between the spatial dimension and the frequency domain dimension in a specific direction, and its expression is: ; ; in, CBR is the feature fusion unit, and the mapping unit is the mapping unit. This indicates channel-dimensional splicing, with AFP being the axial frequency domain projection unit. (4) Finally, attention weights in the vertical and horizontal directions are generated. and Weight and After channel-weighting the input features, they are further refined through a spatial attention module, and finally output through residual connections. : .

[0026] Through the aforementioned mechanism, the BAA module provides the network with crucial, direction-aware local geometric cues, improving the initial localization capability of camouflaged targets.

[0027] Preferably, in step S2, although the BAA module can extract valuable local directional cues from the frequency domain, how to effectively and adaptively feed back the frequency domain information after global context understanding to spatial features to correct the spatial feature distribution and focus on the target region during the complex decoding process remains a challenge. To address this, the present invention designs an Amplitude-Guided Spatial Modulator (AGSM), the structure of which is as follows: Figure 3 As shown.

[0028] The amplitude-guided spatial modulator (AGSM) is configured to perform the following procedure: make Indicates the first The features after the hierarchy is processed by the BAA module, among which These correspond to four levels from deep to shallow; through feature projection and upsampling operations, the transfer of high-level semantic information to low-level features is achieved. ; in, This is a bilinear upsampling operation. This represents the multi-scale features after fusion; For the deepest layer ( ), based solely on cross-layer fusion features As input, output features For other levels ( The input of the AGSM module is determined by the output characteristics of the previous level AGSM module. Cross-layer fusion features with the current layer Assembled by splicing, and through The number of channels is adjusted by convolution, and it is denoted as... : ; First, the Fast Fourier Transform (FFT) is used to... Transform to the frequency domain to obtain The amplitude and phase spectra are obtained. After taking the logarithm of the amplitude spectrum to enhance stability, a lightweight frequency domain attention module (FA) is used to analyze the amplitude spectrum. The FA learns the importance of different frequency components through global average pooling and fully connected layers to generate adaptive modulation weights. The modulated amplitude spectrum is combined with the original phase spectrum and reconstructed by inverse Fourier transform (IFFT) to obtain the frequency domain enhanced features. The process is represented as follows: ; in, Represents grouped convolution. and These represent FFT and IFFT respectively, and FA represents the frequency domain attention module. Represents a unit complex number consisting of the original phase spectrum; Meanwhile, input features Spatial domain features are also extracted through a lightweight convolutional path. A gated fusion unit (GFU) is introduced to adaptively fuse frequency domain enhancement information and spatial structure information, generating channel-dimensional fusion weights based on the input. The final fusion features are: ; The fused features are output through residual connections and upsampling, serving as the refined features of the current level and passed to the next level.

[0029] Preferably, in step S2, during the decoding stage, the network needs to integrate features from different levels, each carrying rich semantic information and fine spatial details. However, simple feature addition or concatenation operations are difficult to adaptively balance the importance of features at different scales, easily leading to inconsistent multi-scale prediction responses and blurred boundaries. Therefore, this invention designs a multi-scale adaptive fusion module (MSAF), the structure of which is as follows: Figure 4 As shown.

[0030] The multi-scale adaptive fusion module MSAF is configured to perform the following process: The module input consists of feature maps at four scales. First, each feature map is projected to a uniform number of channels using a 1×1 convolution and downsampled to the same spatial size. Then, a depthwise separable convolutional enhancement block (EDSB) combined with channel attention (SEBlock) is used to enhance the features at each scale, thereby strengthening their discriminative power and suppressing redundancy, resulting in enhanced features. ; Four enhanced features spliced ​​along the channel dimension This is used to achieve more refined multi-scale feature fusion; Subsequently, the weight generation network WGNet predicts a pixel-level fusion weight for each scale feature at each spatial location. Weighting and learnable temperature coefficient Normalization to achieve finer fusion; the ultimate multi-scale unified representation. It is obtained by weighted summation of the enhancement features at each scale according to their respective weights: ; ; Characterization It also serves for subsequent prediction and monitoring: The features are upsampled and concatenated with the original features from each decoding stage, and then processed by the edge enhancement module. Processing is used to generate detailed prediction maps for each stage. ;at the same time An edge prediction map is generated using a lightweight edge prediction head, EH. And auxiliary supervision is provided by edge ground truth values ​​derived from the real mask: .

[0031] Preferably, step S3 includes end-to-end training and optimization using a deep supervision strategy: Prediction graphs of the four stages of the network decoder All are supervised by the region segmentation loss, which is composed of the binary cross-entropy loss. Compared with the intersection and comparison loss Together constitute; Introducing edge-assisted supervision to enhance boundary details, edge ground truth The edge prediction is obtained by processing the real mask with the Sobel operator and Gaussian filter. Generated by the MSAF module and using Dice loss. Supervision is performed; total loss function Defined as: ; in For the real mask, To balance the weights.

[0032] Preferably, in step S4, structural measurement The calculation method is as follows: ; in, This represents structural similarity based on target perception. This represents region-aware structural similarity. It is a hyperparameter, and its value range is... , The higher the value, the better the detection effect.

[0033] Preferably, in step S4, the alignment metric is enhanced. The calculation method is as follows: ; in, For the prediction map and the ground truth map at the pixel level Alignment balance at the location These are the height and width of the image, respectively; the calculated values ​​are... The higher the value, the better the detection effect.

[0034] Preferably, in step S4, the weighted metric... The calculation method is as follows: ; ; ; in, Indicates weighted precision. Indicates the weighted recall rate. To balance the importance of precision and recall; Represents pixels Spatial weights are used to measure the importance of a pixel in the target structure; , , These represent the true positive, false positive, and false negative cases at the pixel level; the calculated... The higher the value, the more accurate the test result.

[0035] Preferably, in step S4, the mean absolute error The calculation method is as follows: ; in, This represents the predicted probability or normalized pixel value. This represents the corresponding true pixel value. These are the height and width of the image, respectively; the calculated values ​​are... The smaller the value, the smaller the error and the better the detection effect.

[0036] Example 2: To verify the effectiveness of the proposed progressive frequency-spatial-domain collaborative optimization network PFSNet, this embodiment selects a wildlife monitoring scenario for experimentation. In this scenario, target animals such as snow leopards and chameleons often blend seamlessly into their natural environment, exhibiting strong camouflage, making them difficult to accurately identify using traditional detection methods.

[0037] The experiments used four publicly available benchmark datasets for camouflaged target detection: CAMO, COD10K, CHAMELEON, and NC4K. The training set consisted of 1000 images from CAMO and 3040 images from COD10K, totaling 4040 images. The test sets were the complete CHAMELEON dataset (76 images), the complete NC4K dataset (4121 images), and the test sets for CAMO and COD10K (250 and 2026 images respectively).

[0038] The input images are uniformly scaled to a resolution of 512×512, and data augmentation is performed using random flipping, rotation, and cropping.

[0039] The method PFSNet of this invention is compared with current mainstream camouflaged target detection methods, including SINet-V2, PFNet, FEDER, and ZoomNet; the evaluation metric used is a structural metric. Enhance alignment metrics Weighted metrics and The comparison results are shown in Table 1.

[0040] Table 1: Performance comparison of different methods on four datasets;

[0041] From Table 1 and Figure 5 As can be seen, the PFSNet proposed in this invention achieves state-of-the-art results on all evaluation metrics across all four datasets; for example, on the most challenging CAMO dataset, PFSNet achieves the best results. It reached 0.867, a 1.5% improvement over the second-best ZoomNet; It reached 0.924, an increase of 1.3%; The accuracy reached 0.819, an improvement of 1.7%; the MSE decreased to 0.043, with a relative error reduction of 12%; PFSNet also showed significant advantages on large-scale datasets such as COD10K and NC4K. This fully demonstrates the effectiveness of the progressive collaborative mechanism of "directional cue capture - global spectral domain calibration - cross-scale discrimination refinement" constructed in this invention. By capturing directional local structural cues through the band axial attention module BAA, achieving adaptive modulation of the frequency domain to the space through the amplitude guided spatial modulator AGSM, and stabilizing structural consistency through the multi-scale adaptive fusion module MSAF, PFSNet can more accurately locate highly camouflaged targets and recover clear boundary details. Especially in complex scenarios such as wildlife monitoring, this method can significantly reduce missed detections and false detections, and has extremely high practical value.

Claims

1. A progressive camouflage target detection method based on dual-domain fusion, characterized in that, Includes the following steps: S1, Data Preparation and Model Input: Obtain a public dataset, divide it into a training set and a test set, and preprocess the images in the training set; S2, Network Design and Construction: Construct an asymptotic frequency domain-spatial domain co-optimization network, which includes an encoder, a frequency band axial attention module, an amplitude-guided spatial modulator, and a multi-scale adaptive fusion module; The frequency band axial attention module integrates discrete wavelet transform and axial attention to capture directional local structural cues of camouflaged targets; Amplitude-guided spatial modulators introduce a learnable frequency-domain attention mechanism to adaptively reweight the amplitude spectrum and modulate frequency-domain information into spatial features; The multi-scale adaptive fusion module fuses multi-scale features through pixel-level adaptive weighted fusion and introduces an edge-assisted supervision branch; S3, Model Optimization: The preprocessed training set images are input into the progressive frequency-spatial co-optimization network. The network is trained and optimized end-to-end using a deep supervision strategy and a total loss function. The total loss function combines region segmentation loss and edge-assisted supervision loss. S4, Model Performance Evaluation: The performance of the trained network is evaluated using test set images, employing metrics including structural measures. Enhance alignment metrics Weighted metrics and The evaluation metrics, including those used in the study, are employed to verify the effectiveness of camouflage target detection.

2. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S1, the public dataset uses four public benchmark datasets: CAMO, COD10K, CHAMELEON, and NC4K; the training set consists of partial images from CAMO and partial images from COD10K; the test set consists of the complete CHAMELEON dataset, the complete NC4K dataset, and the test sets of CAMO and COD10K.

3. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S2, the Bandwidth Axial Attention Module (BAA) is configured to execute the following process: (1) For input features, the BAA module is at the deepest level Its input is only the feature map of the current level, while in other levels, it takes the BAA output feature map of the previous level. and the output characteristics of the current level SMT backbone network The inputs from two levels are fused using a gated fusion unit (GFU). The process is as follows: ； ； in, For element-wise multiplication, 3 3 convolutions, For batch normalization, express Activation function; (2) The BAA module performs two parallel processes simultaneously: Spatial axial features were extracted along the vertical and horizontal directions using adaptive average pooling AAP. and : ； Two-dimensional discrete wavelet transform is performed on the input features to extract frequency domain subbands, and then the axial frequency domain projection unit (AFP) is used to... and Subbands are processed to obtain frequency domain cues representing the vertical and horizontal directions, respectively. and ; (3) The spatial axial features and their corresponding frequency domain cues are spliced ​​together in the channel dimension and fused through the feature fusion unit CBR to generate fused features. and This models the correlation between the spatial dimension and the frequency domain dimension in a specific direction, and its expression is: ； ； in, CBR is the feature fusion unit, and the mapping unit is the mapping unit. This indicates channel-dimensional splicing, with AFP being the axial frequency domain projection unit. (4) Finally, attention weights in the vertical and horizontal directions are generated. and Weight and After channel-weighting the input features, they are further refined through a spatial attention module, and finally output through residual connections. : 。 4. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S2, the amplitude-guided spatial modulator AGSM is configured to perform the following procedure: make Indicates the first The features after the hierarchy is processed by the BAA module, among which These correspond to four levels from deep to shallow; through feature projection and upsampling operations, the transfer of high-level semantic information to low-level features is achieved. ； in, This is a bilinear upsampling operation. This represents the multi-scale features after fusion; For the deepest layer ( ), based solely on cross-layer fusion features As input, output features For other levels ( The input of the AGSM module is determined by the output characteristics of the previous level AGSM module. Cross-layer fusion features with the current layer Assembled by splicing, and through The number of channels is adjusted by convolution, and it is denoted as... : ； First, the Fast Fourier Transform (FFT) is used to... Transform to the frequency domain to obtain The amplitude and phase spectra are obtained. After taking the logarithm of the amplitude spectrum to enhance stability, a lightweight frequency domain attention module (FA) is used to analyze the amplitude spectrum. The FA learns the importance of different frequency components through global average pooling and fully connected layers to generate adaptive modulation weights. The modulated amplitude spectrum is combined with the original phase spectrum and reconstructed by inverse Fourier transform (IFFT) to obtain the frequency domain enhanced features. The process is represented as follows: ； in, Represents grouped convolution. and These represent FFT and IFFT respectively, and FA represents the frequency domain attention module. Represents a unit complex number consisting of the original phase spectrum; Meanwhile, input features Spatial domain features are also extracted through a lightweight convolutional path. A gated fusion unit (GFU) is introduced to adaptively fuse frequency domain enhancement information and spatial structure information, generating channel-dimensional fusion weights based on the input. The final fusion features are: ； The fused features are output through residual connections and upsampling, serving as the refined features of the current level and passed to the next level.

5. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S2, the multi-scale adaptive fusion module MSAF is configured to perform the following process: The module input consists of feature maps at four scales. First, each feature map is projected to a uniform number of channels using a 1×1 convolution and downsampled to the same spatial size. Then, a depthwise separable convolutional enhancement block (EDSB) combined with channel attention (SEBlock) is used to enhance the features at each scale, thereby strengthening their discriminative power and suppressing redundancy, resulting in enhanced features. ; Four enhanced features spliced ​​along the channel dimension This is used to achieve more refined multi-scale feature fusion; Subsequently, the weight generation network WGNet predicts a pixel-level fusion weight for each scale feature at each spatial location. Weighting and learnable temperature coefficient Normalization to achieve finer fusion; the ultimate multi-scale unified representation. It is obtained by weighted summation of the enhancement features at each scale according to their respective weights: ； ； Characterization It also serves for subsequent prediction and monitoring: The features are upsampled and concatenated with the original features from each decoding stage, and then processed by the edge enhancement module. Processing is used to generate detailed prediction maps for each stage. ;at the same time An edge prediction map is generated using a lightweight edge prediction head, EH. And auxiliary supervision is provided by edge ground truth values ​​derived from the real mask: 。 6. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, Step S3 includes end-to-end training and optimization using a deep supervision strategy: Prediction graphs of the four stages of the network decoder All are supervised by the region segmentation loss, which is composed of the binary cross-entropy loss. Compared with the intersection and comparison loss Together constitute; Introducing edge-assisted supervision to enhance boundary details, edge ground truth The edge prediction is obtained by processing the real mask with the Sobel operator and Gaussian filter. Generated by the MSAF module and using Dice loss. Supervision is performed; total loss function Defined as: ； in For the real mask, To balance the weights.

7. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S4, structural measurement The calculation method is as follows: ； in, This represents structural similarity based on target perception. This represents region-aware structural similarity. It is a hyperparameter, and its value range is... , The higher the value, the better the detection effect.

8. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S4, the alignment metric is enhanced. The calculation method is as follows: ； in, For the prediction map and the ground truth map at the pixel level Alignment balance at the location These are the height and width of the image, respectively; the calculated values ​​are... The higher the value, the better the detection effect.

9. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S4, weighted measurement The calculation method is as follows: ； ； ； in, Indicates weighted precision. Indicates the weighted recall rate. To balance the importance of precision and recall; Represents pixels Spatial weights are used to measure the importance of a pixel in the target structure; , , These represent the true positive, false positive, and false negative cases at the pixel level; the calculated... The higher the value, the more accurate the test result.

10. The progressive camouflage target detection method based on dual-domain fusion according to claim 1, characterized in that, In step S4, the mean absolute error The calculation method is as follows: ； in, This represents the predicted probability or normalized pixel value. This represents the corresponding true pixel value. These are the height and width of the image, respectively; the calculated values ​​are... The smaller the value, the smaller the error and the better the detection effect.

Images