Lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance

Abstract

The application discloses a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, and is implemented according to the following steps: step 1, data preprocessing and lightweight coding; step 2, constructing a multi-stage Haar wavelet frequency prior branch; step 3, constructing a wavelet-based feature transformation module; step 4, constructing a frequency-guided fusion module; and step 5, constructing an asymmetric multi-branch decoding and prediction module. In the feature extraction and fusion process, the layer normalization, residual connection and feedforward neural network are introduced into the overall network structure, which not only improves the stability in the deep model training process, alleviates the gradient vanishing and gradient explosion problems, but also enhances the preservation ability of original feature information and the overall feature representation ability.

Description

Technical Field

[0001] This invention belongs to the technical field of computer vision and image segmentation methods, specifically relating to a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance. Background Technology

[0002] Camouflaged target detection aims to identify and segment targets that are highly similar to their surroundings from complex backgrounds. Because camouflaged targets have a high degree of consistency with the background in terms of color, texture, brightness, and boundary morphology, conventional salient target detection or semantic segmentation methods struggle to reliably extract effective target cues, easily leading to problems such as missing target regions, blurred boundaries, and false detections of the background.

[0003] Existing deep learning methods typically improve detection performance through complex backbone networks, attention mechanisms, or multi-scale fusion structures. However, many of these methods rely on a large number of parameters and significant computational overhead, resulting in inefficiencies in real-time detection, mobile deployment, and resource-constrained scenarios. Furthermore, when the spatial contrast between the target and background is weakened by camouflage, relying solely on spatial texture and semantic features is insufficient to adequately distinguish the target from the interfering background. Summary of the Invention

[0004] The purpose of this invention is to provide a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, which solves the problems of weak difference between camouflage targets and background and unclear boundaries in the prior art.

[0005] The technical solution adopted in this invention is a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, which is implemented according to the following steps: Step 1: Data preprocessing and lightweight coding; Step 2: Construct multi-level Haar wavelet frequency prior branches; Step 3: Construct a wavelet-based feature transformation module; Step 4: Construct the frequency-guided fusion module; Step 5: Construct an asymmetric multi-branch decoding and prediction module.

[0006] The invention is further characterized in that, Step 1 is implemented in the following steps: Step 1.1: Collect natural scene images with pixel-level camouflage target annotations. The images include animal camouflage, artificial camouflage, complex texture backgrounds and low-contrast boundary scenes. The images are divided into training set, validation set and test set. The input images are then subjected to random flipping, scaling, cropping and color perturbation data augmentation operations. Step 1.2: Adjust the input RGB image to a uniform size and perform normalization processing by channel to obtain the normalized input image I. The calculation formula is shown below:

[0007] In the formula, and Let be the mean and standard deviation of channel c on the training set, respectively. This represents the original input RGB image. This represents the normalized input image; Step 1.3: Input the normalized image into the lightweight encoder and extract the spatial features of the four stages according to the hierarchy. The calculation formula is shown in the following formula:

[0008] Where E(·) represents the lightweight encoder, and Fk represents the spatial characteristics of the output of the k-th stage; Step 1.4: Perform 1×1 convolutional channel compression on the features of each stage. The calculation formula is shown in the following formula:

[0009] In the formula, Represents the spatial characteristics of the k-th stage Features after 1×1 convolution compression; Step 2 is implemented in the following steps: Step 2.1: Perform a three-level two-dimensional Haar wavelet decomposition on each color channel of the input image to obtain low-frequency approximate components and high-frequency detail components in multiple directions. The calculation formula is shown in the following equation:

[0010] Where X represents any single-channel image signal of the input image I, wavedec2() represents the two-dimensional wavelet decomposition function, cA3 is the third-level low-frequency approximation component, and cHl, cVl and cDl are the first-level horizontal, vertical and diagonal high-frequency detail components, respectively. Step 2.2: Stack the high-frequency components in the horizontal, vertical and diagonal directions of each level according to the RGB channels to construct a multi-level high-frequency prior. Step 2.3: Stack the third-level low-frequency approximation components according to the RGB channels to construct the low-frequency structure prior; Step 2.4: Scale the high-frequency prior and low-frequency prior to a uniform resolution and expand to a uniform number of channels. The calculation formula is shown below: (5) Step 3 is implemented in the following steps: Step 3.1: Input the compressed stage features into the wavelet-based feature transformation module to perform Haar discrete wavelet transform on the feature domain information. The calculation formula is shown below:

[0011] Where LL represents the low-frequency subband, and LH, HL, and HH represent high-frequency detail subbands in different directions, respectively; Step 3.2: Perform lightweight convolution projection on the low-frequency and high-frequency components respectively, as shown in the following formula:

[0012] in, Indicates low-frequency projection characteristics. Indicates high-frequency projection characteristics; Step 3.3: Reconstruct the low-frequency structural response and high-frequency detail response, and output frequency enhancement features through edge enhancement and residual connection. The calculation formula is shown in the following formula: (8) Where Z represents the intermediate feature after recombination of low-frequency structural response and high-frequency detail response; Step 3.4: Perform edge enhancement and upsampling on the recombined features, and perform residual connection with the input features. The calculation formula is shown in the following formula: (9) Where Z′ represents the feature after edge enhancement is performed on the recombined feature Z. This represents the feature enhanced by the wavelet feature transformation module in the k-th stage; Step 4 is implemented in the following steps: Step 4.1: Channel alignment is performed on the frequency enhancement features, high-frequency priors, and low-frequency priors, respectively. The calculation formula is shown in the following formula:

[0013] in, This represents the spatial features after alignment with the frequency-enhanced features. This represents the features aligned with high-frequency priors. This represents the features aligned with the low-frequency priors; Step 4.2: Add the aligned spatial features, high-frequency priors, and low-frequency priors element-wise to obtain the initial fused feature f. The calculation formula is shown below:

[0014] Step 4.3: Generate high-resolution branch features using a lightweight high- and low-resolution interaction operator. and low-resolution branch features The calculation formula is shown below:

[0015] Step 4.4: Perform depthwise convolution and upsampling on the low-resolution branch, perform edge enhancement on the high-resolution branch, and then concatenate and reassemble to obtain the fused features. The calculation formula is shown below:

[0016] Here, EA represents the edge enhancement module, CBG represents the combination of convolution, batch normalization and GELU activation, and LiteOctConv is used to complete the information exchange of high and low resolution features with low computational cost.

[0017] The lightweight resolution interaction operator includes a high-resolution branch and a low-resolution branch. The high-resolution branch is used to preserve target edges and texture details, while the low-resolution branch is used to aggregate a stable context structure.

[0018] Step 5 is implemented in the following steps: Step 5.1: Input the fused features into the asymmetric multi-branch decoding module and perform channel projection using a 1×1 convolution. The calculation formula is shown below:

[0019] Where X represents the fusion feature of the input asymmetric multi-branch decoding module, This represents the features projected after a 1×1 convolution channel; Step 5.2: Use 3×3, 1×3, and 3×1 depthwise separable convolutional branches to extract local structural information in different directions, and obtain the weighted aggregated features. The calculation formula is shown below:

[0020] in, , , This represents the lightweight branch weights corresponding to the three convolutional branches; Step 5.3: Perform batch normalization, GELU activation, and efficient channel attention on the branch aggregation results, and output the residual refinement features. The calculation formula is shown in the following formula:

[0021] Step 5.4: Generate a coarse prediction from the deepest features, and then refine the shallow features step by step using a coarse-to-fine approach. The calculation formula is shown in the following equation:

[0022] Where F represents the feature map to be guided, and P represents the prediction map from the previous stage. This indicates that the gradient operation is stopped. This represents the Sigmoid activation function; Step 5.5: Calculate the structural loss and edge loss based on the multi-scale prediction results, and train the final camouflaged target detection model. The calculation formula is shown below: (19) Where L represents the total training loss of the model. Indicates structural loss. Let Y represent the predicted map at the k-th scale, and let Y represent the pixel-level annotation map of the real camouflaged target. This represents the edge loss weighting coefficient. This represents the Sobel gradient magnitude operator; The final output includes a camouflaged target probability map, a binary segmentation mask, and a contour overlay result, which are used to locate hidden targets in complex backgrounds at the pixel level.

[0023] The beneficial effects of this invention are: This invention presents a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance. It explicitly introduces low-frequency structure and high-frequency detail priors through multi-level Haar wavelet decomposition, enhancing the separability of camouflage targets in weak spatial contrast scenarios. A wavelet-based feature transformation module reconstructs the frequency response in the feature domain, improving the representation of target boundaries and texture details. A frequency-guided fusion module achieves efficient interaction between spatial features and frequency priors, suppressing background interference. An asymmetric multi-branch decoding module completes coarse-to-fine prediction with low computational cost, improving the completeness and real-time performance of camouflage target segmentation. Attached Figure Description

[0024] Figure 1 This is the overall network architecture diagram of the lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance of the present invention; Figure 2 This is a structural diagram of the wavelet-based feature transformation module in the lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance of the present invention. Figure 3 This is a diagram of the frequency-guided fusion module and LiteOctConv structure in the lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance of this invention. Figure 4 This is a structural diagram of the asymmetric multi-branch decoding module in the lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0026] This invention presents a lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance, the network architecture of which is shown in the figure below. Figure 1 As shown, please follow these steps: Step 1: Data preprocessing and lightweight coding; Step 2: Construct multi-level Haar wavelet frequency prior branches; Step 3: Construct a wavelet-based feature transformation module; Step 4: Construct the frequency-guided fusion module; Step 5: Construct an asymmetric multi-branch decoding and prediction module.

[0027] Example 1 A lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance, wherein step 1 is implemented according to the following steps: Step 1.1: Collect natural scene images with pixel-level camouflage target annotations. The images include animal camouflage, artificial camouflage, complex texture backgrounds and low-contrast boundary scenes. The images are divided into training set, validation set and test set. The input images are then subjected to random flipping, scale transformation, cropping and color perturbation data augmentation operations to improve the network's ability to adapt to changes in target pose, scale and brightness. Step 1.2: Adjust the input RGB image to a uniform size and perform normalization processing by channel to obtain the normalized input image I. The calculation formula is shown below:

[0028] In the formula, and Let be the mean and standard deviation of channel c on the training set, respectively. This represents the original input RGB image. This represents the normalized input image; Step 1.3: Input the normalized image into the lightweight encoder and extract the spatial features of the four stages according to the hierarchy. The calculation formula is shown in the following formula:

[0029] Where E(·) represents the lightweight encoder, and Fk represents the spatial characteristics of the output of the k-th stage; Step 1.4: Perform 1×1 convolutional channel compression on the features of each stage. The calculation formula is shown in the following formula:

[0030] In the formula, Represents the spatial characteristics of the k-th stage Features after 1×1 convolution compression; Example 2 A lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance, wherein step 2 is implemented according to the following steps: Step 2.1: Perform a three-level two-dimensional Haar wavelet decomposition on each color channel of the input image to obtain low-frequency approximate components and high-frequency detail components in multiple directions. The calculation formula is shown in the following equation:

[0031] Where X represents any single-channel image signal of the input image I, wavedec2() represents the two-dimensional wavelet decomposition function, cA3 is the third-level low-frequency approximation component, and cHl, cVl and cDl are the first-level horizontal, vertical and diagonal high-frequency detail components, respectively. Step 2.2: Stack the high-frequency components of each level in the horizontal, vertical and diagonal directions according to the RGB channels to construct a multi-level high-frequency prior, which is used to describe the boundary, fine texture and local mutation information of the camouflaged target. Step 2.3: Stack the third-level low-frequency approximation components according to the RGB channels to construct a low-frequency structural prior, which is used to describe the target body region and stable contour structure; Step 2.4: Scale the high-frequency prior and low-frequency prior to a uniform resolution and expand to a uniform number of channels. The calculation formula is shown below: (5) Example 3 A lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, wherein the structure diagram of the wavelet-based feature transformation module is as follows: Figure 2 As shown, step 3 is implemented in the following steps: Step 3.1: Input the compressed stage features into the wavelet-based feature transformation module to perform Haar discrete wavelet transform on the feature domain information. The calculation formula is shown below:

[0032] Where LL represents the low-frequency subband, and LH, HL, and HH represent high-frequency detail subbands in different directions, respectively; Step 3.2: Perform lightweight convolution projection on the low-frequency and high-frequency components respectively, as shown in the following formula:

[0033] in, Indicates low-frequency projection characteristics. Indicates high-frequency projection characteristics; Step 3.3: The low-frequency structural response and high-frequency detail response are recombined, and the nonlinear expression is enhanced by the GELU function. The calculation formula is shown below: (8) Where Z represents the intermediate feature after recombination of low-frequency structural response and high-frequency detail response; Step 3.4: Perform edge enhancement and upsampling on the recombined features, and perform residual connection with the input features. The calculation formula is shown in the following formula: (9) Where Z′ represents the feature after edge enhancement is performed on the recombined feature Z. This represents the feature enhanced by the wavelet feature transformation module in the k-th stage; Example 4 A lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance is proposed, wherein the frequency-guided fusion module and LiteOctConv structure diagram are shown below. Figure 3 As shown, step 4 is implemented in the following steps: Step 4.1: Channel alignment is performed on the frequency enhancement features, high-frequency priors, and low-frequency priors, respectively. The calculation formula is shown in the following formula:

[0034] in, This represents the spatial features after alignment with the frequency-enhanced features. This represents the features aligned with high-frequency priors. This represents the features aligned with the low-frequency priors; Step 4.2: Add the aligned spatial features, high-frequency priors, and low-frequency priors element-wise to obtain the initial fused features. The calculation formula is shown below:

[0035] Step 4.3: Employ a lightweight high-low resolution interaction operator to define high-resolution branch features. and low-resolution branch features The calculation formula is shown below:

[0036] Step 4.4: Perform depthwise convolution and upsampling on the low-resolution branch, perform edge enhancement on the high-resolution branch, and then concatenate and reassemble to obtain the fused features. The calculation formula is shown below:

[0037] Here, EA represents the edge enhancement module, CBG represents the combination of convolution, batch normalization and GELU activation, and LiteOctConv is used to complete the information exchange of high and low resolution features with low computational cost.

[0038] The lightweight high- and low-resolution interaction operator includes a high-resolution branch and a low-resolution branch. The high-resolution branch is used to preserve the target edge and texture details, while the low-resolution branch is used to aggregate a stable context structure.

[0039] Example 5 A lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, with an asymmetric multi-branch decoding module structure diagram as shown below. Figure 4 As shown, step 5 is implemented in the following steps: Step 5.1: Input the fused features into the asymmetric multi-branch decoding module and perform channel projection using a 1×1 convolution. The calculation formula is shown below:

[0040] Where X represents the fusion feature of the input asymmetric multi-branch decoding module, This represents the features projected after a 1×1 convolution channel; Step 5.2: Use 3×3, 1×3, and 3×1 depthwise separable convolutional branches to extract local structural information in different directions, and obtain the weighted aggregated features. The calculation formula is shown below:

[0041] in, , , This represents the lightweight branch weights corresponding to the three convolutional branches; Step 5.3: Perform batch normalization, GELU activation, and efficient channel attention on the branch aggregation results, and output the residual refinement features. The calculation formula is shown in the following formula:

[0042] Step 5.4: Generate a coarse prediction from the deepest features, and then refine the shallow features step by step using a coarse-to-fine approach. The calculation formula is shown in the following equation:

[0043] Where F represents the feature map to be guided, and P represents the prediction map from the previous stage. This indicates that the gradient operation is stopped. This represents the Sigmoid activation function; Step 5.5: Calculate the structural loss and edge loss based on the multi-scale prediction results, and train the final camouflaged target detection model. The calculation formula is shown below: (19) Where L represents the total training loss of the model. Indicates structural loss. Let Y represent the predicted map at the k-th scale, and let Y represent the pixel-level annotation map of the real camouflaged target. This represents the edge loss weighting coefficient. This represents the Sobel gradient magnitude operator; Example 6 Table 1 shows a comparison of the network model trained using the method of this invention with the results of various camouflaged target detection methods: Table 1

[0044] In the table, : Structural metrics. : Average enhanced alignment metric. : Weighted F-measure. Mean absolute error. Param: Number of parameters. FPS: Frames per second. The best score is indicated in bold.

[0045] As can be seen from Table 1, the present invention achieves high weighted F values ​​on multiple public test sets while maintaining a high inference speed, indicating that the frequency-guided recombination structure can effectively enhance the detection performance of camouflaged targets.

[0046] This invention presents a lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance. In the feature extraction and fusion process, layer normalization, residual connections, and feedforward neural networks are introduced into the overall network structure. This not only improves the stability of deep model training and alleviates the gradient vanishing and gradient exploding problems, but also enhances the ability to retain original feature information and the overall feature representation ability. It solves the problems of weak difference between camouflage targets and background, unclear boundaries, large model computation, and insufficient utilization of frequency information in existing technologies.

Claims

1. A lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance, characterized in that, The specific steps are as follows: Step 1: Data preprocessing and lightweight coding; Step 2: Construct multi-level Haar wavelet frequency prior branches; Step 3: Construct a wavelet-based feature transformation module; Step 4: Construct the frequency-guided fusion module; Step 5: Construct an asymmetric multi-branch decoding and prediction module.

2. The lightweight camouflage target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, Step 1 is implemented in the following steps: Step 1.1: Collect natural scene images with pixel-level camouflage target annotations. The images include animal camouflage, artificial camouflage, complex texture backgrounds and low-contrast boundary scenes. The images are divided into training set, validation set and test set. The input images are then subjected to random flipping, scaling, cropping and color perturbation data augmentation operations. Step 1.2: Adjust the input RGB image to a uniform size and perform normalization processing by channel to obtain the normalized input image I. The calculation formula is shown below: In the formula, and Let be the mean and standard deviation of channel c on the training set, respectively. This represents the original input RGB image. This represents the normalized input image; Step 1.3: Input the normalized image into the lightweight encoder and extract the spatial features of the four stages according to the hierarchy. The calculation formula is shown in the following formula: Where E(·) represents the lightweight encoder, and Fk represents the spatial characteristics of the output of the k-th stage; Step 1.4: Perform 1×1 convolutional channel compression on the features of each stage. The calculation formula is shown in the following formula: In the formula, Represents the spatial characteristics of the k-th stage Features after 1×1 convolution compression.

3. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, Step 2 is implemented in the following steps: Step 2.1: Perform a three-level two-dimensional Haar wavelet decomposition on each color channel of the input image to obtain low-frequency approximate components and high-frequency detail components in multiple directions. The calculation formula is shown in the following equation: Where X represents any single-channel image signal of the input image I, wavedec2() represents the two-dimensional wavelet decomposition function, cA3 is the third-level low-frequency approximation component, and cHl, cVl and cDl are the first-level horizontal, vertical and diagonal high-frequency detail components, respectively. Step 2.2: Stack the high-frequency components in the horizontal, vertical and diagonal directions of each level according to the RGB channels to construct a multi-level high-frequency prior. Step 2.3: Stack the third-level low-frequency approximation components according to the RGB channels to construct the low-frequency structure prior; Step 2.4: Scale the high-frequency prior and low-frequency prior to a uniform resolution and expand to a uniform number of channels. The calculation formula is shown below: （5）。 4. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Input the compressed stage features into the wavelet-based feature transformation module to perform Haar discrete wavelet transform on the feature domain information. The calculation formula is shown below: Where LL represents the low-frequency subband, and LH, HL, and HH represent high-frequency detail subbands in different directions, respectively; Step 3.2: Perform lightweight convolution projection on the low-frequency and high-frequency components respectively, as shown in the following formula: in, Indicates low-frequency projection characteristics. Indicates high-frequency projection characteristics; Step 3.3: Reconstruct the low-frequency structural response and high-frequency detail response, and output frequency enhancement features through edge enhancement and residual connection. The calculation formula is shown in the following formula: （8） Where Z represents the intermediate feature after recombination of low-frequency structural response and high-frequency detail response; Step 3.4: Perform edge enhancement and upsampling on the recombined features, and perform residual connection with the input features. The calculation formula is shown in the following formula: （9） Where Z′ represents the feature after edge enhancement is performed on the recombined feature Z. This represents the feature enhanced by the wavelet feature transformation module in the k-th stage.

5. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, Step 4 is implemented in the following steps: Step 4.1: Channel alignment is performed on the frequency enhancement features, high-frequency priors, and low-frequency priors, respectively. The calculation formula is shown in the following formula: in, This represents the spatial features after alignment with the frequency-enhanced features. This represents the features aligned with high-frequency priors. This represents the features aligned with the low-frequency priors; Step 4.2: Add the aligned spatial features, high-frequency priors, and low-frequency priors element-wise to obtain the initial fused features. The calculation formula is shown below: Step 4.3: Generate high-resolution branch features using a lightweight high- and low-resolution interaction operator. and low-resolution branch features The calculation formula is shown below: Step 4.4: Perform depthwise convolution and upsampling on the low-resolution branch, perform edge enhancement on the high-resolution branch, and then concatenate and reassemble to obtain the fused features. The calculation formula is shown below: Here, EA represents the edge enhancement module, CBG represents the combination of convolution, batch normalization and GELU activation, and LiteOctConv is used to complete the information exchange of high and low resolution features with low computational cost.

6. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 5, characterized in that, The lightweight high- and low-resolution interaction operator includes a high-resolution branch and a low-resolution branch. The high-resolution branch is used to preserve target edges and texture details, while the low-resolution branch is used to aggregate a stable context structure.

7. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, Step 5 is implemented in the following steps: Step 5.1: Input the fused features into the asymmetric multi-branch decoding module and perform channel projection using a 1×1 convolution. The calculation formula is shown below: Where X represents the fusion feature of the input asymmetric multi-branch decoding module, This represents the features projected after a 1×1 convolution channel; Step 5.2: Use 3×3, 1×3, and 3×1 depthwise separable convolutional branches to extract local structural information in different directions, and obtain the weighted aggregated features. The calculation formula is shown below: in, , , This represents the lightweight branch weights corresponding to the three convolutional branches; Step 5.3: Perform batch normalization, GELU activation, and efficient channel attention on the branch aggregation results, and output the residual refinement features. The calculation formula is shown in the following formula: Step 5.4: Generate a coarse prediction from the deepest features, and then refine the shallow features step by step using a coarse-to-fine approach. The calculation formula is shown in the following equation: Where F represents the feature map to be guided, and P represents the prediction map from the previous stage. This indicates that the gradient operation is stopped. This represents the Sigmoid activation function; Step 5.5: Calculate the structural loss and edge loss based on the multi-scale prediction results, and train the final camouflaged target detection model. The calculation formula is shown below: （19） Where L represents the total training loss of the model. Indicates structural loss. Let Y represent the predicted map at the k-th scale, and let Y represent the pixel-level annotation map of the real camouflaged target. This represents the edge loss weighting coefficient. This represents the Sobel gradient magnitude operator.

8. The lightweight camouflaged target detection method based on hierarchical wavelet decomposition and frequency guidance according to claim 1, characterized in that, The final output includes a camouflaged target probability map, a binary segmentation mask, and a contour overlay result, which are used to locate hidden targets in complex backgrounds at the pixel level.

Images