A camouflage animal detection method based on frequency domain decoupling and edge information fusion

Abstract

The application provides a camouflage animal detection method based on frequency domain decoupling and edge information fusion, and relates to the technical field of image processing. The application uses an image frequency domain decoupler to dynamically separate low-frequency (global structure) and high-frequency (edge details) features through a learnable Gaussian filter group, and efficiently utilizes the features in an edge contour predictor and a camouflage target detector, thereby effectively improving the detection accuracy of camouflage animals. In high-frequency edge blur scenes such as dry leaf butterflies, the high-frequency channel enhances the target contour response, and the missing detection rate is significantly reduced. The low-frequency channel retains the overall shape of the target, and improves the recognition robustness of large-scale camouflage targets such as zebra groups. At the same time, in interference environments such as rain, fog and night, the frequency domain features are less affected by sensor noise. By replacing multi-modal data with frequency domain decoupling, the detection performance has high detection accuracy and stability.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, in particular to a camouflaged animal detection method based on frequency domain decoupling and edge information fusion. Background Art

[0002] In nature, organisms adapt to their surroundings by employing disruptive color and shadow removal to create camouflage, thereby protecting themselves from predators. Effectively identifying camouflage and finding targets not only helps us understand the human perceptual system but also has practical applications in polyp segmentation, military countermeasures, and endangered species detection. Common methods for detecting camouflaged animals primarily analyze texture, color, optical flow, and multimodal data.

[0003] Currently, a method for detecting camouflaged animals based on the YOLO object detection network is available. By improving the YOLO network structure or replacing the feature extractor, it focuses on spatial features such as texture and color. For example, the receptive field of the convolutional layer is adjusted to enhance the capture of local details, or an attention mechanism is introduced to increase the weight of the target region. However, because this method relies solely on single spatial features, when the camouflaged animal and the background have similar texture / color (such as a dead leaf butterfly among fallen leaves), the spatial features cannot distinguish the target edge, resulting in a sharp drop in detection accuracy against complex backgrounds. Currently, a multimodal method for detecting camouflaged animals and plants in complex environments is also available. This method integrates depth information with image features, leveraging the additional cues contained in depth information to improve detection performance. However, this method requires additional sensor equipment such as depth cameras, which is costly and complex to deploy. Furthermore, in extreme scenarios (such as rain and fog interfering with depth sensing or high infrared noise at night), the quality of cross-modal data degrades, the fusion performance degrades, and detection accuracy also decreases. Finally, each modality requires its own feature extractor and feature fusion processor, which typically requires high-performance computing equipment, increasing the complexity and cost of the method. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to propose a camouflaged animal detection method based on frequency domain decoupling and edge information fusion, comprising:

[0005] Step 1: Obtain a camouflaged target image of an animal from an existing dataset, and obtain the true annotation of the camouflaged target image, where the area where the animal is located in the true annotation is white, and the area other than the animal in the camouflaged target image is black;

[0006] Step 2: Input the camouflaged target image into the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain camouflaged animal detection results, wherein the camouflaged animal detection results include a first camouflaged animal prediction map, a second camouflaged animal prediction map, a third camouflaged animal prediction map, and a fourth camouflaged animal prediction map; the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model includes an image feature extractor, an image frequency domain decoupler, an edge contour predictor, and a camouflaged animal detector;

[0007] Step 3: Based on the first camouflaged animal prediction map, the second camouflaged animal prediction map, the third camouflaged animal prediction map, the fourth camouflaged animal prediction map, the true annotation and the edge contour prediction map, the total loss function value is calculated, and based on the total loss function value, the parameters in the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model are iteratively updated to obtain the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model;

[0008] Step 4: Obtain the image to be detected and input it into the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain four camouflaged animal prediction images. The camouflaged animal prediction image with the highest resolution among the four camouflaged animal prediction images is used as the final detection result.

[0009] Optionally, step 2 specifically includes:

[0010] Step 2.1: Inputting the disguised target image into an image feature extractor to obtain a first feature map, a second feature map, a third feature map, and a fourth feature map, wherein the sizes of the first feature map, the second feature map, the third feature map, and the fourth feature map are all different;

[0011] Step 2.2: Input the first feature map, the second feature map, the third feature map, and the fourth feature map into the image frequency domain decoupler to obtain a low-frequency feature map and a high-frequency feature map corresponding to the first feature map, a low-frequency feature map and a high-frequency feature map corresponding to the second feature map, a low-frequency feature map and a high-frequency feature map corresponding to the third feature map, and a low-frequency feature map and a high-frequency feature map corresponding to the fourth feature map;

[0012] Step 2.3: Input the first feature map and the high-frequency feature map corresponding to the first feature map, the second feature map and the high-frequency feature map corresponding to the second feature map, the third feature map and the high-frequency feature map corresponding to the third feature map, and the fourth feature map and the high-frequency feature map corresponding to the fourth feature map into the edge contour predictor to obtain an edge contour prediction map;

[0013] Step 2.4: Input the first feature map and the low-frequency feature map corresponding to the first feature map, the second feature map and the low-frequency feature map corresponding to the second feature map, the third feature map and the corresponding low-frequency feature map, the fourth feature map and the corresponding low-frequency feature map, and the edge contour prediction map into the camouflaged animal detector to obtain a first camouflaged animal prediction map, a second camouflaged animal prediction map, a third camouflaged animal prediction map, and a fourth camouflaged animal prediction map.

[0014] Optionally, step 2.2 specifically includes:

[0015] Step 2.2.1: Take the first feature map, or the second feature map, or the third feature map, or the fourth feature map as the first current feature map, and perform a two-dimensional fast Fourier transform on the first current feature map to obtain a transformed feature map. F ( μ , v ), which is specifically expressed by the following formula:

[0016] F ( μ , v )=FFT{f( x , y )};

[0017] Among them, f( x , y ) is the first current feature map in the two-dimensional space coordinate ( x , y ), FFT stands for two-dimensional fast Fourier transform, μ Represents the frequency component index in the horizontal direction of the frequency domain, v Indicates the frequency component index in the vertical direction of the frequency domain;

[0018] Step 2.2.2: Pass the transformed feature map through a Gaussian low-pass filter to obtain low-frequency domain features L ( μ , v ), which is specifically achieved through the following formula:

[0019] ;

[0020] in, D ( μ , v ) is the distance from the frequency coordinate to the origin, σ L represents the trainable parameters of the low-frequency pathway;

[0021] The transformed feature map is passed through a Gaussian high-pass filter to obtain high-frequency domain features H ( μ , v), which is specifically achieved through the following formula:

[0022] ;

[0023] in, σ H represents the trainable parameters of the high-frequency pathway;

[0024] Step 2.2.3: Low frequency domain features L ( μ , v ) and high-frequency features H ( μ , v ) perform inverse Fourier transform respectively to obtain the low-frequency feature map f L ( x , y ) and high-frequency feature maps f H ( x , y ), which is specifically expressed by the following formula:

[0025] ;

[0026] ;

[0027] Among them, FFT -1 represents the inverse Fourier transform;

[0028] Then, the low-frequency feature map and the high-frequency feature map corresponding to the first feature map, the low-frequency feature map and the high-frequency feature map corresponding to the second feature map, the low-frequency feature map and the high-frequency feature map corresponding to the third feature map, and the low-frequency feature map and the high-frequency feature map corresponding to the fourth feature map are obtained.

[0029] Optionally, step 2.3 specifically includes:

[0030] Step 2.3.1: Perform a high-frequency feature fusion operation on the fourth feature map and the high-frequency feature map corresponding to the fourth feature map to obtain a first high-frequency fused feature map;

[0031] Step 2.3.2: Perform a bidirectional positioning operation on the edge information of the first high-frequency fusion feature map to obtain a first contour map;

[0032] Step 2.3.3: Fuse the first contour map with the third feature map to obtain a first contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the high-frequency feature maps corresponding to the first contour fusion feature map and the third feature map to obtain a second contour map.

[0033] Step 2.3.4: Fuse the second contour map with the second feature map to obtain a second contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the second contour fusion feature map and the high-frequency feature map corresponding to the second feature map to obtain a third contour map.

[0034] Step 2.3.5: Fuse the third contour map with the first feature map to obtain a third contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the third contour fusion feature map and the high-frequency feature map corresponding to the first feature map to obtain a fourth contour map.

[0035] Step 2.3.6: Perform convolution on the fourth contour map to obtain the edge contour prediction map.

[0036] Optionally, step 2.3.1 specifically includes:

[0037] The high-frequency feature map corresponding to the fourth feature map is subjected to 1x1 convolution to obtain the first convolution feature map, the first convolution feature map and the fourth feature map are spliced ​​in the channel dimension to obtain the first spliced ​​feature map, the first spliced ​​feature map is processed by the channel attention mechanism to obtain the first spatial weight map, the first spatial weight map is feature added with the fourth feature map to obtain the first high-frequency fusion feature map.

[0038] Optionally, step 2.3.2 specifically includes:

[0039] Perform 1x1 convolution on the first high-frequency fusion feature map to obtain a second convolution feature map, process the second convolution feature map through the foreground module to obtain a target foreground feature map, specifically, process the second convolution feature map respectively through three 3x3 convolutions to obtain a first intermediate feature map, a second intermediate feature map and a third intermediate feature map, perform dilated convolution with a dilation rate of 1 on the first intermediate feature map to obtain a first foreground feature map, add features of the first foreground feature map and the second intermediate feature map to obtain a fourth intermediate feature map, perform dilated convolution with a dilation rate of 2 on the fourth intermediate feature map to obtain a second foreground feature map, add features of the second foreground feature map and the third intermediate feature map to obtain a fifth intermediate feature map, perform dilated convolution with a dilation rate of 3 on the fifth intermediate feature map to obtain a third foreground feature map, perform channel-level splicing on the first foreground feature map, perform 3x3 convolution and normalization on the spliced ​​feature map to obtain a target foreground feature map;

[0040] The second convolution feature map is processed by a background module to obtain a target background feature map. The background module has the same operation as the foreground module, but the weight matrix of the dilated convolution and the convolution is different;

[0041] The first contour map is obtained by element-wise multiplication of the target foreground feature map and the target background feature map.

[0042] Optionally, step 2.4 specifically includes:

[0043] Step 2.4.1: Process the fourth feature map and the low-frequency feature map corresponding to the fourth feature map using information enhancement, low-frequency feature fusion, information filtering, and information localization to obtain a first localization feature map. The first localization feature map and the edge contour prediction map are concatenated in the channel dimension and subjected to channel attention, convolution, and upsampling to obtain a first camouflaged animal prediction map.

[0044] Step 2.4.2: Process the third feature map and the low-frequency feature map corresponding to the third feature map using information enhancement, low-frequency feature fusion, information filtering, and information positioning to obtain a second positioning feature map. Use information correction to process the second positioning feature map and the first camouflaged animal prediction map to obtain a first information-corrected feature map. Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and perform channel attention, convolution, and upsampling to obtain a second camouflaged animal prediction map.

[0045] Step 2.4.3: Fuse the first information correction feature map with the second feature map to obtain a first fused feature map. Based on information enhancement, low-frequency feature fusion operations, information filtering, and information positioning, process the low-frequency feature maps corresponding to the first fused feature map and the second feature map to obtain a third positioning feature map. Based on the second camouflaged animal prediction map, perform information correction on the third positioning feature map to obtain a second information correction feature map. Concatenate the second information correction feature map with the edge contour prediction map in the channel dimension and perform channel attention, convolution, and upsampling to obtain a third camouflaged animal prediction map.

[0046] Step 2.4.4: Fuse the second information correction feature map with the first feature map to obtain a second fused feature map. Based on information enhancement, low-frequency feature fusion operation, information filtering and information positioning, process the second fused feature map and the low-frequency feature map corresponding to the first feature map to obtain a fourth positioning feature map. Based on the third camouflaged animal prediction map, perform information correction on the fourth positioning feature map to obtain a third information correction feature map. The third information correction feature map and the edge contour prediction map are spliced ​​in the channel dimension and subjected to channel attention, convolution and upsampling to obtain the fourth camouflaged animal prediction map.

[0047] Optionally, step 2.4.1 specifically includes:

[0048] Step 2.4.1.1: Perform feature enhancement on the fourth feature map to obtain a first enhanced feature map;

[0049] Specifically, 1 x 1 convolution, first dilated convolution, second dilated convolution, third dilated convolution and global average pooling are performed on the fourth feature map, and the obtained feature maps are spliced. 1 x 1 convolution is performed on the spliced ​​feature map to obtain a first enhanced feature map, wherein the expansion rate of the first dilated convolution, the expansion rate of the second dilated convolution and the expansion rate of the third dilated convolution are all different;

[0050] Step 2.4.1.2: Perform a low-frequency feature fusion operation on the low-frequency feature maps corresponding to the first enhanced feature map and the fourth feature map to obtain a first low-frequency fused feature map;

[0051] Specifically, a 1 x 1 convolution is performed on the low-frequency feature map corresponding to the fourth feature map to obtain a second convolution feature map, the second convolution feature map is spliced ​​with the first enhanced feature map in the channel dimension to obtain a second spliced ​​feature map, the second spliced ​​feature map is processed by the channel attention mechanism to obtain a second spatial weight map, and the second spatial weight map is feature added with the first enhanced feature map to obtain a first low-frequency fusion feature map;

[0052] Step 2.4.1.3: Filter the first low-frequency fusion feature map to obtain a first filtered feature map;

[0053] Specifically, the first low-frequency fusion feature map is subjected to 1x1 convolution, channel reorganization, and channel grouping to obtain m group features, the m group features are subjected to channel maximum pooling to obtain the maximum response of the intra-group feature, and the m group features are subjected to channel average pooling to obtain the average response of the intra-group feature. The maximum response of the intra-group feature and the average response of the intra-group feature are spliced ​​in the channel dimension and passed through 3x3 convolution to obtain a convolved feature map, and the convolved feature map is feature-added with the first low-frequency fusion feature map to obtain a first filtered feature map;

[0054] Step 2.4.1.4: Perform information positioning on the first filtering feature map to obtain a first positioning feature map;

[0055] Specifically, the first filtered feature map is subjected to 3x3 convolution, normalization and ReLU activation function to obtain the first precise positioning feature map; the first filtered feature map is subjected to normalization and multi-head conversion attention to obtain the global positioning feature map;

[0056] The first filter feature map is split to obtain a first split feature map and a second split feature map, where the size of the first split feature map is different from the size of the second split feature map, and the first split feature map is subjected to 3x3 convolution, normalization and Relu activation function to obtain a first local feature map. At the same time, the first split feature map is subjected to 3x3 convolution, normalization and Relu activation function to obtain a second local feature map, where the parameters of the two 3x3 convolutions of the first split feature map are different, the first local feature map and the second local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a second precise positioning feature map;

[0057] The second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a third local feature map. At the same time, the second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a fourth local feature map, wherein the parameters of the two 5x5 convolutions of the second split feature map are different; the third local feature map and the fourth local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a third precise positioning feature map;

[0058] Add the first precise positioning feature map, the second precise positioning feature map, the third precise positioning feature map, and the global positioning feature map, and perform 3x3 convolution on the added feature map to obtain a global-local feature map. Add the global-local feature map and the first filtered feature map to obtain a first positioning feature map.

[0059] Step 2.4.1.5: The first positioning feature map and the edge contour prediction map are spliced ​​in the channel dimension and subjected to channel attention, convolution and upsampling to obtain the first camouflaged animal prediction map.

[0060] Optionally, step 2.4.2 specifically includes:

[0061] Step 2.4.2.1: Process the third feature map and the low-frequency feature map corresponding to the third feature map through information enhancement, low-frequency feature fusion operations, information filtering, and information positioning to obtain a second positioning feature map;

[0062] Specifically, the third feature map is enhanced to obtain a second enhanced feature map, a low-frequency feature fusion operation is performed on the second enhanced feature map and the low-frequency feature map corresponding to the third feature map to obtain a second low-frequency fused feature map, information is filtered on the second low-frequency fused feature map to obtain a second filtered feature map, and information positioning is performed on the second filtered feature map to obtain a second positioned feature map;

[0063] Step 2.4.2.2: Based on the first camouflaged animal prediction image, perform information correction and edge contour information fusion on the second positioning feature image to obtain a first information correction feature image;

[0064] Specifically, based on the first camouflaged animal prediction map, the first value UR ( x, y ), which is specifically expressed by the following formula:

[0065] ;

[0066] in, is the two-dimensional coordinate in the first camouflaged animal prediction map ( x , y ) at the eigenvalue;

[0067] Based on the first numerical value, the first camouflaged animal prediction map is divided into an identified area and an uncertain area, wherein the value of the first numerical value corresponding to each coordinate in the identified area is 0, and the value of the first numerical value corresponding to each coordinate in the uncertain area is 1;

[0068] The second positioning feature map is element-wise multiplied with the uncertain area, and the multiplied feature map is input into the residual spatial attention to obtain an unidentified positioning feature map. The second positioning feature map is element-wise multiplied with the identified area, and the multiplied feature map is input into the residual channel attention to obtain an identified positioning feature map. The second positioning feature map is input into the residual spatial attention and the residual channel attention respectively to obtain a spatial attention feature map and a channel attention feature map. The unidentified positioning feature map, the identified positioning feature map, the spatial attention feature map and the channel attention feature map are spliced, and the spliced ​​result is subjected to a 1x1 convolution to obtain a first information correction feature map.

[0069] Step 2.4.2.3: Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and obtain the second camouflaged animal prediction map through channel attention, convolution, and upsampling.

[0070] Optionally, in step 3, the total loss function value is calculated based on the first camouflaged animal prediction map, the second camouflaged animal prediction map, the third camouflaged animal prediction map, the fourth camouflaged animal prediction map, the true annotation, and the edge contour prediction map, which is specifically implemented by the following formula:

[0071] ;

[0072] in, L All Represents the total loss function value, Indicates the i Camouflaged animal prediction map, G Indicates the true markup, Indicates the i Segmentation loss between the camouflaged animal predictions and the ground-truth annotations, Indicates the i The area overlap loss between the camouflaged animal prediction map and the true annotation, Represents the edge contour prediction map and the true edge The edge alignment loss between To camouflage the outline of the animal in the target image;

[0073] Among them, the parameters in the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model are iteratively updated, including the parameters in the image feature extractor, the trainable parameters of the high-frequency path and the trainable parameters of the low-frequency path in the image frequency domain decoupler, the parameters of the convolution in the edge contour predictor, and the parameters of the convolution in the camouflaged animal detector and the parameters of the multi-head conversion attention.

[0074] The beneficial effects of adopting the above technical solution are:

[0075] This invention uses an image frequency-domain decoupler to dynamically separate low-frequency (global structure) and high-frequency (edge ​​detail) features via a learnable Gaussian filter bank. These features are then efficiently utilized in an edge contour predictor and camouflaged target detector, effectively improving the detection accuracy of camouflaged animals. In scenes with high-frequency, blurred edges, such as dead leaf butterflies, the high-frequency pathway enhances the target contour response, significantly reducing missed detection rates. The low-frequency pathway preserves the overall target shape, improving the robustness of recognition for large-scale camouflaged targets such as herds of zebras. Furthermore, because frequency-domain features are less susceptible to sensor noise in interfering environments such as rain, fog, and at night, this method, through frequency-domain decoupling instead of multimodal data, achieves highly accurate and stable detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 Schematic diagram of a flow chart of a camouflaged animal detection method based on frequency domain decoupling and edge information fusion in an embodiment of the present invention;

[0077] Figure 2 Schematic diagram of the processing flow of the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model in an embodiment of the present invention;

[0078] Figure 3 Schematic diagram of the process of high-frequency feature fusion operation in an embodiment of the present invention;

[0079] Figure 4 Schematic diagram of the process of edge information bidirectional positioning operation in an embodiment of the present invention;

[0080] Figure 5 A schematic diagram of the information enhancement process in an embodiment of the present invention;

[0081] Figure 6 Schematic diagram of the process of low-frequency feature fusion operation in an embodiment of the present invention;

[0082] Figure 7 Schematic diagram of the information filtering process in an embodiment of the present invention;

[0083] Figure 8 A schematic diagram of the process of information positioning in an embodiment of the present invention;

[0084] Figure 9 A schematic diagram of a flow chart of information correction in an embodiment of the present invention;

[0085] Figure 10 Schematic diagram of experimental results in an embodiment of the present invention, wherein (a) is the image to be detected of animal 1, (a1) is the true labeling of animal 1, (a2) is the final detection result of animal 1, (b) is the image to be detected of animal 2, (b1) is the true labeling of animal 2, (b2) is the final detection result of animal 2, (c) is the image to be detected of animal 3, (c1) is the true labeling of animal 3, (c2) is the final detection result of animal 3, (d) is the image to be detected of animal 4, (d1) is the true labeling of animal 4, (d2) is the final detection result of animal 4, (e) is the image to be detected of animal 4, (e1) is the true labeling of animal 4, and (e2) is the final detection result of animal 4. DETAILED DESCRIPTION

[0086] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0087] In view of the problems existing in the prior art, the present invention provides a camouflaged animal detection method based on frequency domain decoupling and edge information fusion, combined with Figure 1 , which may include the following steps:

[0088] Step 1: Obtain a camouflaged target image of an animal from an existing dataset, and obtain the true annotation of the camouflaged target image, where the area where the animal is located in the true annotation is white, and the area other than the animal in the camouflaged target image is black;

[0089] Existing datasets include CAMO (Camouflaged Object Detection Dataset), COD10K (Camouflaged Object Detection Benchmark), and NC4K (Extended Camouflage Detection Dataset). These three datasets comprise multi-dimensional feature libraries of 15,371 images. The CAMO dataset provides 1,250 carefully annotated images of camouflaged objects, covering eight categories of mimicry, including insects and reptiles. The COD10K dataset contributes 10,000 cross-species samples, grouped by five lighting conditions (day / dusk / night / fog / rain) and three levels of background complexity (simple / medium / complex). Each sample is accompanied by an HSV color space distribution histogram. The NC4K dataset supplements this with 4,121 high-resolution images (1920×1080 to 4096×2160). Its multimodal annotation system includes pixel-level instance segmentation, material reflectance vectors, and biomimetic motion trajectory simulation data.

[0090] Using a stratified-temporal hybrid partitioning method, the total sample size was divided into a training set (65%, 9,991 images), a validation set (20%, 3,074 images), and a test set (15%, 2,306 images) based on the proportion of the source datasets. The training set consists of the first 65% of CAMO time series data (812 images), 6,500 random images from COD10K, and the first 65% of NC4K samples (2,679 images). The validation set consists of the last 20% of the time series data from the latest released versions of each dataset. The test set strictly uses the official hidden test subsets of each benchmark library to maintain objectivity. All partitioning processes are verified by SHA-256 hashing to ensure sample independence.

[0091] Step 2: Input the camouflaged target image into the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain the camouflaged animal detection results, which include the first camouflaged animal prediction map, the second camouflaged animal prediction map, the third camouflaged animal prediction map and the fourth camouflaged animal prediction map; the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model includes an image feature extractor, an image frequency domain decoupler, an edge contour predictor and a camouflaged animal detector; combined Figure 2 The specific steps include:

[0092] Step 2.1: Inputting the disguised target image into an image feature extractor to obtain a first feature map, a second feature map, a third feature map, and a fourth feature map, wherein the sizes of the first feature map, the second feature map, the third feature map, and the fourth feature map are all different;

[0093] Specifically, the present invention uses the PvTv2-B4 network as the image feature extractor. In the specific implementation process, the size of the four feature maps obtained is the resolution of the disguised target image. 、 、 、 .

[0094] Step 2.2: Input the first feature map, the second feature map, the third feature map, and the fourth feature map into the image frequency domain decoupler to obtain a low-frequency feature map and a high-frequency feature map corresponding to the first feature map, a low-frequency feature map and a high-frequency feature map corresponding to the second feature map, a low-frequency feature map and a high-frequency feature map corresponding to the third feature map, and a low-frequency feature map and a high-frequency feature map corresponding to the fourth feature map;

[0095] The image frequency domain decoupler is based on frequency domain analysis theory and achieves frequency domain decoupling of image features through a learnable filter bank. The core innovation lies in converting the spatial domain feature map to the frequency domain and then using adaptive low-pass / high-pass filters to separate the frequency components, i.e., frequency domain decoupling. Finally, the feature information of different frequencies is input into the edge contour predictor and camouflaged animal detector, respectively, to help FDEF-Net better generate the edge contours of camouflaged animals and improve detection accuracy. The specific implementation includes the following steps:

[0096] Step 2.2.1: Take the first feature map, or the second feature map, or the third feature map, or the fourth feature map as the first current feature map, and perform a two-dimensional fast Fourier transform on the first current feature map to obtain a transformed feature map. F ( μ , v ), which is specifically expressed by the following formula:

[0097] F ( μ , v )=FFT{f( x , y )};

[0098] Among them, f( x , y ) is the first current feature map in the two-dimensional space coordinate ( x , y ), FFT stands for two-dimensional fast Fourier transform, μ Represents the frequency component index in the horizontal direction of the frequency domain, v Indicates the frequency component index in the vertical direction of the frequency domain;

[0099] Step 2.2.2: Pass the transformed feature map through a Gaussian low-pass filter to obtain low-frequency domain features L ( μ ,v ), which is specifically achieved through the following formula:

[0100] ;

[0101] in, D ( μ , v ) is the distance from the frequency coordinate to the origin, σ L represents the trainable parameters of the low-frequency pathway;

[0102] The transformed feature map is passed through a Gaussian high-pass filter to obtain high-frequency domain features H ( μ , v ), which is specifically achieved through the following formula:

[0103] ;

[0104] in, σ H represents the trainable parameters of the high-frequency pathway;

[0105] Step 2.2.3: Low frequency domain features L ( μ , v ) and high-frequency features H ( μ , v ) perform inverse Fourier transform respectively to obtain the low-frequency feature map f L ( x , y ) and high-frequency feature maps f H ( x , y ), which is specifically expressed by the following formula:

[0106] ;

[0107] ;

[0108] Among them, FFT -1 represents the inverse Fourier transform;

[0109] The first feature map, the second feature map, the third feature map and the fourth feature map are respectively processed by the image frequency domain decoupler to obtain the low-frequency feature map and the high-frequency feature map corresponding to the first feature map, the low-frequency feature map and the high-frequency feature map corresponding to the second feature map, the low-frequency feature map and the high-frequency feature map corresponding to the third feature map, and the low-frequency feature map and the high-frequency feature map corresponding to the fourth feature map.

[0110] Step 2.3: Input the first feature map and the high-frequency feature map corresponding to the first feature map, the second feature map and the high-frequency feature map corresponding to the second feature map, the third feature map and the high-frequency feature map corresponding to the third feature map, and the fourth feature map and the high-frequency feature map corresponding to the fourth feature map into the edge contour predictor to obtain an edge contour prediction map;

[0111] The edge contour predictor primarily processes the output features of the image feature extractor and the image frequency domain decoupler, and uses this processed information to predict the outlines of camouflaged animals in the image, thereby providing guidance for subsequent camouflaged animal detection based on edge contour information. The edge contour predictor primarily involves two operations: high-frequency feature information fusion and bidirectional edge information localization. This is achieved through the following steps:

[0112] Step 2.3.1: Perform a high-frequency feature fusion operation on the fourth feature map and the high-frequency feature map corresponding to the fourth feature map to obtain a first high-frequency fused feature map;

[0113] Combine Figure 3 , the high-frequency feature map corresponding to the fourth feature map (i.e., the input feature map) is subjected to 1x1 convolution to obtain the first convolution feature map, the first convolution feature map and the fourth feature map are spliced ​​in the channel dimension to obtain the first spliced ​​feature map, the first spliced ​​feature map is processed by the channel attention mechanism to obtain the first spatial weight map, the first spatial weight map is feature-added with the fourth feature map to obtain the first high-frequency fusion feature map (i.e., the output feature map).

[0114] Step 2.3.2: Perform a bidirectional positioning operation on the edge information of the first high-frequency fusion feature map to obtain a first contour map;

[0115] Edge-based bidirectional localization primarily consists of a foreground module and a background module. The foreground module processes foreground information containing the animal, while the background module processes background information surrounding the animal. The intersection of foreground and background information is the characteristic information of the target's edge contour. Edge-based bidirectional localization not only mines foreground and background information to enhance target positioning, but also generates edge contour predictions, providing edge guidance for camouflaged animal detection.

[0116] Specific, combined Figure 4, perform 1x1 convolution on the first high-frequency fusion feature map (i.e., the input feature map) to obtain a second convolution feature map, process the second convolution feature map through the foreground module to obtain a target foreground feature map, specifically, process the second convolution feature map through three 3x3 convolutions respectively to obtain a first intermediate feature map, a second intermediate feature map, and a third intermediate feature map, perform dilated convolution with a dilation rate of 1 on the first intermediate feature map to obtain a first foreground feature map, add the features of the first foreground feature map and the second intermediate feature map to obtain a fourth intermediate feature map, perform dilated convolution with a dilation rate of 2 on the fourth intermediate feature map to obtain a second foreground feature map, add the features of the second foreground feature map and the third intermediate feature map to obtain a fifth intermediate feature map, perform dilated convolution with a dilation rate of 3 on the fifth intermediate feature map to obtain a third foreground feature map, perform channel-level splicing on the first foreground feature map, perform 3x3 convolution and normalization on the spliced ​​feature map to obtain a target foreground feature map;

[0117] The second convolution feature map is processed by a background module to obtain a target background feature map. The background module has the same operation as the foreground module, but the weight matrices of the dilated convolution and the convolution are different.

[0118] That is to say, combined Figure 4 , perform 1x1 convolution on the second convolution feature map, then perform 3x3 convolution, dilated convolution and channel-level splicing, perform 3x3 convolution and normalization on the spliced ​​feature map to obtain the target background feature map.

[0119] The target foreground feature map and the target background feature map are element-wise multiplied to obtain the first contour map (i.e., the output feature map).

[0120] Step 2.3.3: Fuse the first contour map with the third feature map to obtain a first contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the high-frequency feature maps corresponding to the first contour fusion feature map and the third feature map to obtain a second contour map.

[0121] Among them, the first contour fusion feature map is Figure 3 The input feature map in the third feature map corresponds to the high-frequency feature map: Figure 3 The high-frequency feature map corresponding to the input feature map in , Figure 3 The output feature map is Figure 4 The input feature map of Figure 4 The output feature map is the second contour map.

[0122] Step 2.3.4: Fuse the second contour map with the second feature map to obtain a second contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the second contour fusion feature map and the high-frequency feature map corresponding to the second feature map to obtain a third contour map.

[0123] Among them, the second contour fusion feature map is Figure 3 The input feature map in the second feature map corresponds to the high-frequency feature map: Figure 3 The high-frequency feature map corresponding to the input feature map in , Figure 3 The output feature map is Figure 4 The input feature map of Figure 4 The output feature map is the third contour map.

[0124] Step 2.3.5: Fuse the third contour map with the first feature map to obtain a third contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the third contour fusion feature map and the high-frequency feature map corresponding to the first feature map to obtain a fourth contour map.

[0125] Among them, the third contour fusion feature map is Figure 3 The input feature map in the first feature map corresponds to the high-frequency feature map: Figure 3 The high-frequency feature map corresponding to the input feature map in , Figure 3 The output feature map is Figure 4 The input feature map of Figure 4 The output feature map is the fourth contour map.

[0126] It should be noted that the high-frequency feature fusion operation in steps 2.3.3 to 2.3.5 is the same as the high-frequency feature fusion operation in step 2.3.1, except that the input feature map and the output feature map are different. The edge information bidirectional positioning operation in steps 2.3.3 to 2.3.5 is the same as the high-frequency feature fusion operation in step 2.3.2, except that the input feature map and the output feature map are different.

[0127] Step 2.3.6: Perform convolution on the fourth contour map to obtain the edge contour prediction map.

[0128] Step 2.4: Input the first feature map and the low-frequency feature map corresponding to the first feature map, the second feature map and the low-frequency feature map corresponding to the second feature map, the third feature map and the corresponding low-frequency feature map, the fourth feature map and the corresponding low-frequency feature map, and the edge contour prediction map into the camouflaged animal detector to obtain a first camouflaged animal prediction map, a second camouflaged animal prediction map, a third camouflaged animal prediction map, and a fourth camouflaged animal prediction map.

[0129] The camouflaged animal detector primarily processes the outputs of the image feature extractor, the image frequency domain decoupler, and the edge contour predictor, and then generates a detection result based on the processed information. The camouflaged animal detector primarily includes operations such as information enhancement, information filtering, information localization, information correction, and the fusion of low-frequency information and edge contour information.

[0130] Step 2.4.1: Process the fourth feature map and the low-frequency feature map corresponding to the fourth feature map using information enhancement, low-frequency feature fusion, information filtering, and information localization to obtain a first localization feature map. The first localization feature map and the edge contour prediction map are concatenated in the channel dimension and subjected to channel attention, convolution, and upsampling to obtain a first camouflaged animal prediction map.

[0131] Step 2.4.1.1: Perform feature enhancement on the fourth feature map to obtain a first enhanced feature map;

[0132] Information enhancement mainly uses the ASPP structure to process different feature maps and enhance them. Its main principle is to capture contextual information through multi-scale hole convolution to improve image segmentation accuracy. Its workflow is divided into four steps: 1) The input feature maps (i.e., the first feature map, the second feature map, the third feature map, and the fourth feature map) are sent to multiple parallel branches. Each branch uses dilated convolution with different dilation rates (such as 6, 12, and 18) to obtain local features of different receptive fields while maintaining resolution; 2) A global average pooling branch is added to extract the global context; 3) All branch outputs are spliced ​​and fused with 1x1 convolution; 4) The output is a feature map that integrates local details, multi-scale information, and global semantics, so that the model can simultaneously recognize targets of different sizes.

[0133] Specific, combined Figure 5 , the fourth feature map (i.e., the input feature map) is respectively subjected to 1 x 1 convolution, the first dilated convolution, the second dilated convolution, the third dilated convolution and global average pooling, and the obtained feature maps are spliced, and the spliced ​​feature maps are subjected to 1 x 1 convolution to obtain the first enhanced feature map (i.e., the output feature map), wherein the expansion rate of the first dilated convolution, the expansion rate of the second dilated convolution and the expansion rate of the third dilated convolution are all different. In the specific implementation process, the expansion rate of the first dilated convolution is r=6, the expansion rate of the second dilated convolution is r=12, and the expansion rate of the third dilated convolution is r=18.

[0134] Step 2.4.1.2: Perform a low-frequency feature fusion operation on the low-frequency feature maps corresponding to the first enhanced feature map and the fourth feature map to obtain a first low-frequency fused feature map;

[0135] The low-frequency information fusion operation is essentially the same as the high-frequency feature information fusion operation. The enhanced feature map and the corresponding low-frequency feature map are received as dual inputs, and the high-frequency fused feature map is output. The low-frequency features are first processed through a 1x1 convolution to compress the channel dimension, and then concatenated with the enhanced feature map in the channel dimension. The concatenated features are then processed using an attention mechanism to generate a spatial weight map, dynamically allocating the fusion ratio between the enhanced feature map and the low-frequency features. The weighted fusion result is then added to the original enhanced feature map via a residual connection.

[0136] Specific, combined Figure 6 , perform 1 x 1 convolution on the low-frequency feature map corresponding to the fourth feature map (i.e., the low-frequency feature map corresponding to the input feature map) to obtain a second convolution feature map, concatenate the second convolution feature map with the first enhanced feature map (i.e., the input feature map) in the channel dimension to obtain a second concatenated feature map, process the second concatenated feature map through the channel attention mechanism to obtain a second spatial weight map, perform feature addition on the second spatial weight map and the first enhanced feature map to obtain a first low-frequency fusion feature map (i.e., the output feature map);

[0137] Since feature information will bring about information redundancy after information enhancement and fusion of low-frequency feature information, the information filtering operation is mainly to reduce the generation of redundant information while ensuring that sufficient information is obtained, thereby reducing the impact of redundant information on detection and reducing the computational burden.

[0138] Step 2.4.1.3: Filter the first low-frequency fusion feature map to obtain a first filtered feature map;

[0139] Among them, information filtering is mainly divided into two stages: feature preparation and feature filtering. In the feature preparation stage, the input feature map is expanded to n times the original number of channels through 1x1 convolution. After that, the channel reorganization operation ( ) The feature maps after expanding the number of channels are reorganized and arranged in the channel dimension. Finally, through the grouping operation ( ) Divide the channel reorganized features into m group features ( In the feature filtering stage, the m group features will be processed through channel maximum pooling and channel average pooling to generate the maximum response of the feature within the group. The average corresponding to the characteristics within the group These maximum and average responses are then concatenated in the channel dimension and fused via a 3x3 convolution. Finally, the fused features are added to the original input feature map to form the final output.

[0140] Specific, combined Figure 7, perform 1x1 convolution, channel reorganization, and channel grouping on the first low-frequency fusion feature map (i.e., input feature map) to obtain m group features, perform channel maximum pooling on the m group features to obtain the maximum feature response within the group, and perform channel average pooling on the m group features to obtain the average feature response within the group, concatenate the maximum feature response within the group and the average feature response within the group in the channel dimension and perform 3x3 convolution to obtain the convolved feature map, perform feature addition on the convolved feature map and the first low-frequency fusion feature map to obtain the first filtered feature map (i.e., output feature map);

[0141] Step 2.4.1.4: Perform information positioning on the first filtering feature map to obtain a first positioning feature map;

[0142] After information filtering, these representative features are deeply analyzed through information positioning operations to provide the necessary information for locating camouflaged animals.

[0143] Specifically, the first filtered feature map is subjected to 3x3 convolution, normalization and ReLU activation function to obtain the first precise positioning feature map; the first filtered feature map is subjected to normalization and multi-head conversion attention to obtain the global positioning feature map;

[0144] Specific, combined Figure 8 , split the first filter feature map (i.e., the input feature map) to obtain a first split feature map and a second split feature map, where the size of the first split feature map is different from the size of the second split feature map, and the first split feature map is subjected to 3x3 convolution, normalization, and Relu activation function to obtain a first local feature map. At the same time, the first split feature map is subjected to 3x3 convolution, normalization, and Relu activation function to obtain a second local feature map, where the parameters of the two 3x3 convolutions of the first split feature map are different, the first local feature map and the second local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a second precise positioning feature map;

[0145] The second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a third local feature map. At the same time, the second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a fourth local feature map, wherein the parameters of the two 5x5 convolutions of the second split feature map are different; the third local feature map and the fourth local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a third precise positioning feature map;

[0146] The first precise positioning feature map, the second precise positioning feature map, the third precise positioning feature map and the global positioning feature map are added together, and the added feature map is subjected to a 3x3 convolution to obtain a global-local feature map. The global-local feature map and the first filtered feature map are added together to obtain a first positioning feature map (i.e., the output feature map).

[0147] Step 2.4.1.5: The first positioning feature map and the edge contour prediction map are spliced ​​in the channel dimension and subjected to channel attention, convolution and upsampling to obtain the first camouflaged animal prediction map.

[0148] Step 2.4.2: Process the third feature map and the low-frequency feature map corresponding to the third feature map using information enhancement, low-frequency feature fusion, information filtering, and information positioning to obtain a second positioning feature map. Use information correction to process the second positioning feature map and the first camouflaged animal prediction map to obtain a first information-corrected feature map. Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and perform channel attention, convolution, and upsampling to obtain a second camouflaged animal prediction map.

[0149] Among them, information enhancement, low-frequency feature fusion operation, information filtering, and information positioning are the same as those in step 2.4.1, except that the input and output feature maps are different.

[0150] Step 2.4.2.1: Process the third feature map and the low-frequency feature map corresponding to the third feature map through information enhancement, low-frequency feature fusion operations, information filtering, and information positioning to obtain a second positioning feature map;

[0151] Specifically, the third feature map is enhanced to obtain a second enhanced feature map, a low-frequency feature fusion operation is performed on the second enhanced feature map and the low-frequency feature map corresponding to the third feature map to obtain a second low-frequency fused feature map, information is filtered on the second low-frequency fused feature map to obtain a second filtered feature map, and information positioning is performed on the second filtered feature map to obtain a second positioned feature map;

[0152] Combine Figure 5 、 Figure 6 、 Figure 7 and Figure 8 To explain, Figure 5 The input feature map is the third feature map, Figure 5 The output feature map of is the second enhanced feature map, Figure 6 The input feature map is the second enhanced feature map, Figure 6 The low-frequency feature map corresponding to the input feature map is the low-frequency feature map corresponding to the third feature map. Figure 6 The output feature map is the second low-frequency fusion feature map, Figure 7 The input feature map is the second low-frequency fusion feature map, Figure 7 The output feature map of is the second filtered feature map, Figure 8 The input feature map is the second filtered feature map, Figure 8 The output feature map of is the second positioning feature map.

[0153] Step 2.4.2.2: Based on the first camouflaged animal prediction image, perform information correction and edge contour information fusion on the second positioning feature image to obtain a first information-corrected feature image. The information correction process primarily utilizes accurate information to resolve uncertainties in the inference process.

[0154] Specific, combined Figure 9 , based on the first camouflaged animal prediction map (i.e., input prediction map), calculate the first value UR( x,y ), which is specifically expressed by the following formula:

[0155] ;

[0156] in, is the two-dimensional coordinate in the first camouflaged animal prediction map ( x , y ) at the eigenvalue;

[0157] Based on the first numerical value, the first camouflaged animal prediction map is divided into an identified area and an uncertain area, wherein the value of the first numerical value corresponding to each coordinate in the identified area is 0, and the value of the first numerical value corresponding to each coordinate in the uncertain area is 1;

[0158] The second positioning feature map (i.e., input feature map) is element-wise multiplied with the uncertain area, and the multiplied feature map is input into the residual spatial attention to obtain an unrecognized positioning feature map. The second positioning feature map is element-wise multiplied with the recognized area, and the multiplied feature map is input into the residual channel attention to obtain an recognized positioning feature map. The second positioning feature map is input into the residual spatial attention and the residual channel attention respectively to obtain a spatial attention feature map and a channel attention feature map. The unrecognized positioning feature map, the recognized positioning feature map, the spatial attention feature map, and the channel attention feature map are spliced ​​together, and the spliced ​​result is subjected to a 1x1 convolution to obtain a first information correction feature map (i.e., output feature map).

[0159] Step 2.4.2.3: Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and obtain the second camouflaged animal prediction map through channel attention, convolution, and upsampling.

[0160] Step 2.4.3: Fuse the first information correction feature map with the second feature map to obtain a first fused feature map, process the low-frequency feature maps corresponding to the first fused feature map and the second feature map based on information enhancement, low-frequency feature fusion operation, information filtering and information positioning to obtain a third positioning feature map, perform information correction on the third positioning feature map based on the second camouflage animal prediction map to obtain a second information correction feature map, and splice the second information correction feature map and the edge contour prediction map in the channel dimension and perform channel attention, convolution and upsampling to obtain a third camouflage animal prediction map.

[0161] In combination Figure 9 , it is explained that Figure 9 The input feature map of the step 2.4.2 is the third positioning feature map, the input prediction map is the second camouflage animal prediction map, and the output feature map is the second information correction feature map.

[0162] Step 2.4.4: Fuse the second information correction feature map with the first feature map to obtain a second fused feature map, process the low-frequency feature maps corresponding to the second fused feature map and the first feature map based on information enhancement, low-frequency feature fusion operation, information filtering and information positioning to obtain a fourth positioning feature map, perform information correction on the fourth positioning feature map based on the third camouflage animal prediction map to obtain a third information correction feature map, and splice the third information correction feature map and the edge contour prediction map in the channel dimension and perform channel attention, convolution and upsampling to obtain a fourth camouflage animal prediction map.

[0163] In combination Figure 9 , it is explained that Figure 9 The input feature map of the step 2.4.2 is the third positioning feature map, the input prediction map is the second camouflage animal prediction map, and the output feature map is the second information correction feature map.

[0164] In the step 2.4.3 and the step 2.4.4, the information enhancement, the low-frequency feature fusion operation, the information filtering and the information positioning are the same as the processing in the step 2.4.1, except that the input and output feature maps are different. The information correction is the same as the processing in the step 2.4.2, except that the input and output feature maps are different.

[0165] Step 3: Based on the first camouflage animal prediction map, the second camouflage animal prediction map, the third camouflage animal prediction map, the fourth camouflage animal prediction map, the real label and the edge contour prediction map, a total loss function value is calculated, and the total loss function value is specifically realized through the following formula:

[0166] ;

[0167] Among them, L All The total loss function value is represented by Ltotal, The firsti Camouflaged animal prediction map, G Indicates the true markup, Indicates the i Segmentation loss between the camouflaged animal predictions and the ground-truth annotations, Indicates the i The area overlap loss between the camouflaged animal prediction map and the true annotation, Represents the edge contour prediction map and the true edge The edge alignment loss between To camouflage the outline of the animal in the target image.

[0168] Among them, the region overlap loss is calculated using the intersection-over-union (IOU) ratio, which can increase the weight value of the target pixel under global constraints; the segmentation loss uses the weighted binary cross entropy loss, which can increase the attention to the target pixel under local constraints; the edge alignment loss is calculated using the pixel-level similarity loss, which can help the model strengthen the learning of edge structures.

[0169] Iteratively update the parameters of the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model. Specifically, it iteratively update the parameters of the image feature extractor, the trainable parameters of the high-frequency path and the trainable parameters of the low-frequency path in the image frequency domain decoupler, the parameters of the convolution in the edge contour predictor, and the parameters of the convolution and multi-head switching attention in the camouflaged animal detector to obtain the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model.

[0170] The FDEF-Net model used the AdamW optimizer with a 1e-4 weight decay. Furthermore, the initial learning rate was 1e-4, and the learning rate was scaled down by a factor of 10 at the 30th, 40th, and 50th epochs. The model was trained for 150 epochs using two data augmentation methods: random rotation and random horizontal flipping.

[0171] Step 4: Obtain the image to be detected and input it into the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain four camouflaged animal prediction images. The camouflaged animal prediction image with the highest resolution among the four camouflaged animal prediction images is used as the final detection result.

[0172] Therefore, by processing the image to be detected, we can get Figure 10, where (a) is the image to be detected of animal 1, (a1) is the true label of animal 1, (a2) is the final detection result of animal 1, (b) is the image to be detected of animal 2, (b1) is the true label of animal 2, (b2) is the final detection result of animal 2, (c) is the image to be detected of animal 3, (c1) is the true label of animal 3, (c2) is the final detection result of animal 3, (d) is the image to be detected of animal 4, (d1) is the true label of animal 4, (d2) is the final detection result of animal 4, (e) is the image to be detected of animal 4, (e1) is the true label of animal 4, and (e2) is the final detection result of animal 4.

[0173] Based on the above technical solution, the present invention and the prior art conducted experiments on the COD10K and NC4K datasets. The experimental results are shown in Table 1.

[0174] Table 1 Performance of different methods on COD10K and NC4K datasets;

[0175]

[0176] Among them, the four indicators in Table 1 are structural similarity (Sm), enhanced alignment (αE), Fβ measure and mean absolute error (MAE): Sm measures the consistency between the predicted results and the true value in the overall and local structure. The closer it is to 1, the better the contour preservation; αE evaluates the pixel and context alignment of the binarized image under multiple thresholds, taking into account both coarse and fine level matching; Fβ comprehensively considers precision (Precision) and recall (Recall), and MAE is the average absolute error between the predicted probability map and the binary true value map on all pixels. The smaller the value, the lower the pixel-level deviation.

[0177] As shown in Table 1, on the COD10K dataset, FDEF-Net significantly outperforms the state-of-the-art methods. Compared to the second-ranked BSA-Net, the proposed method's structural similarity (Sm) improved from 0.818 to 0.891 (+8.8%), average enhanced precision (αE) improved from 0.894 to 0.943 (+5.4%), and the Fβ index improved from 0.699 to 0.826 (+18.2%). Furthermore, the mean absolute error (MAE) decreased from 0.034 to 0.020 (-41.2%). These improvements fully demonstrate the effectiveness of the proposed dual frequency domain + edge decoupling strategy in challenging micro-camouflage scenarios.

[0178] FDEF-Net also maintained its leading position on the NC4K dataset. Its Sm, αE, and Fβ reached 0.890, 0.935, and 0.847, respectively, with MAE reduced to 0.032. Compared to BSA-Net, Sm improved by 5.8%, αE by 3.2%, Fβ by 9.8%, and MAE by 33.3%. Compared to six other methods, including CFNet, UJSC, PFNet, SLSR, UGTR, and PreyNet, FDEF-Net achieved significant advantages in all four core metrics and demonstrated stable detection performance under extreme conditions such as varying illumination and cluttered backgrounds.

[0179] In addition to its quantitative improvements, FDEF-Net also offers advantages in practical applications. Relying solely on standard RGB images, it eliminates the need for expensive infrared, depth, or multispectral sensors, significantly reducing system deployment and maintenance costs. Its end-to-end design accelerates inference and engineering implementation, meeting the dual demands of high precision and efficiency for scenarios such as real-time field monitoring and ecological protection.

[0180] In summary, the proposed FDEF-Net achieves breakthroughs in camouflaged animal detection, both in method design and performance. First, in terms of network structure, the invention innovatively introduces a dual-path frequency domain decoupler—a learnable Gaussian / anti-Gaussian filter—that adaptively separates the low-frequency and high-frequency components of the camouflaged target from the complex background in the frequency domain. Furthermore, a dual-path dilated convolution edge contour predictor for foreground / background accurately captures subtle texture differences. The overall architecture mimics the biological visual information processing process, empowering the model with enhanced perception and reasoning capabilities in complex scenes.

[0181] In summary, the image frequency domain decoupler is designed and constructed. Based on the two-way frequency domain decoupling architecture of the learnable Gaussian filter, the dynamic separation of frequency domain features is realized through the parameterized low-pass filter and the inverse Gaussian high-pass filter. The FFT / IFFT is used to realize the lossless conversion of features between the spatial domain and the frequency domain, and the features in different frequency domains after conversion are used to guide the edge contour prediction and camouflage target detection. The edge contour predictor is also designed and constructed. The feature fusion method is designed to fuse high-frequency feature information and corresponding feature information. The foreground / background double-channel dilated convolution is used to construct a multi-scale perception field through hierarchical superposition of dilution rates, and the edge prediction algorithm based on feature intersection is used to realize pixel-level edge positioning. The camouflage animal detector is also designed and constructed. The camouflage animal detector is constructed by simulating the information processing process in the biological visual system, such as information enhancement, information filtering, information positioning, information correction, etc. In this process, the global information of the low-frequency information enhancement feature is introduced, and the edge contour prediction map is introduced to improve the detection performance of the camouflage animal. The deep neural network FDEF-Net model based on frequency domain decoupling and edge information fusion is also designed and constructed. Unlike existing models, the FDEF-Net model does not rely on other sensor devices, but only needs ordinary image acquisition devices as input. Through the efficient cooperation of the image frequency domain decoupler, the image feature extractor, the edge contour predictor and the camouflage animal detector, the detection of camouflage animals in complex environments is realized.

[0182] The method of the present application proposes a camouflage animal detection method based on frequency domain decoupling and edge information fusion. Considering the characteristics of image frequency domain features, an image frequency domain decoupler is designed to obtain low-frequency features reflecting the global structure of the camouflage target and high-frequency features focusing on the edge of the camouflage target, which are used for camouflage animal detection and edge contour prediction respectively. Considering that the edge contour can help the model to better distinguish camouflage animals and the surrounding environment, an edge contour predictor is designed to use double-channel dilated convolution to process high-frequency features and corresponding feature maps for edge positioning, realizing edge contour prediction of camouflage animals. A camouflage animal detector simulating the biological visual system and capable of efficiently fusing frequency domain information and edge contour prediction maps is designed, which realizes camouflage animal detection in complex environments by enhancing, filtering, positioning and correcting information.

[0183] The above description is only the preferred embodiment of the present disclosure and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the application involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features and the technical features disclosed in the embodiments of the present disclosure (but not limited to) with similar functions are replaced with each other to form a technical solution.

Claims

1. A camouflaged animal detection method based on frequency domain decoupling and edge information fusion, characterized in that: include: Step 1: Obtain a camouflaged target image of an animal from an existing dataset, and obtain the true annotation of the camouflaged target image, where the area where the animal is located in the true annotation is white, and the area other than the animal in the camouflaged target image is black; Step 2: Input the camouflaged target image into the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain camouflaged animal detection results, wherein the camouflaged animal detection results include a first camouflaged animal prediction map, a second camouflaged animal prediction map, a third camouflaged animal prediction map, and a fourth camouflaged animal prediction map; the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model includes an image feature extractor, an image frequency domain decoupler, an edge contour predictor, and a camouflaged animal detector; Step 3: Based on the first camouflaged animal prediction map, the second camouflaged animal prediction map, the third camouflaged animal prediction map, the fourth camouflaged animal prediction map, the true annotation and the edge contour prediction map, the total loss function value is calculated, and based on the total loss function value, the parameters in the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model are iteratively updated to obtain the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model; Step 4: Obtain the image to be detected and input it into the trained frequency domain decoupling and edge information fusion deep neural network FDEF-Net model to obtain four camouflaged animal prediction images. The camouflaged animal prediction image with the highest resolution among the four camouflaged animal prediction images is used as the final detection result.

2. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Inputting the disguised target image into an image feature extractor to obtain a first feature map, a second feature map, a third feature map, and a fourth feature map, wherein the sizes of the first feature map, the second feature map, the third feature map, and the fourth feature map are all different; Step 2.2: Input the first feature map, the second feature map, the third feature map, and the fourth feature map into the image frequency domain decoupler to obtain a low-frequency feature map and a high-frequency feature map corresponding to the first feature map, a low-frequency feature map and a high-frequency feature map corresponding to the second feature map, a low-frequency feature map and a high-frequency feature map corresponding to the third feature map, and a low-frequency feature map and a high-frequency feature map corresponding to the fourth feature map; Step 2.3: Input the first feature map and the high-frequency feature map corresponding to the first feature map, the second feature map and the high-frequency feature map corresponding to the second feature map, the third feature map and the high-frequency feature map corresponding to the third feature map, and the fourth feature map and the high-frequency feature map corresponding to the fourth feature map into the edge contour predictor to obtain an edge contour prediction map; Step 2.4: Input the first feature map and the low-frequency feature map corresponding to the first feature map, the second feature map and the low-frequency feature map corresponding to the second feature map, the third feature map and the corresponding low-frequency feature map, the fourth feature map and the corresponding low-frequency feature map, and the edge contour prediction map into the camouflaged animal detector to obtain a first camouflaged animal prediction map, a second camouflaged animal prediction map, a third camouflaged animal prediction map, and a fourth camouflaged animal prediction map.

3. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 2, characterized in that: Step 2.2 specifically includes: Step 2.2.1: Take the first feature map, or the second feature map, or the third feature map, or the fourth feature map as the first current feature map, and perform a two-dimensional fast Fourier transform on the first current feature map to obtain a transformed feature map. F ( μ , v ), which is specifically expressed by the following formula: F ( μ , v )=FFT{f( x , y )}; Among them, f( x , y ) is the first current feature map in the two-dimensional space coordinate ( x , y ), FFT stands for two-dimensional fast Fourier transform, μ Represents the frequency component index in the horizontal direction of the frequency domain, v Indicates the frequency component index in the vertical direction of the frequency domain; Step 2.2.2: Pass the transformed feature map through a Gaussian low-pass filter to obtain low-frequency domain features L ( μ , v ), which is specifically achieved through the following formula: ； in, D ( μ , v ) is the distance from the frequency coordinate to the origin, σ L represents the trainable parameters of the low-frequency pathway; The transformed feature map is passed through a Gaussian high-pass filter to obtain high-frequency domain features H ( μ , v ), which is specifically achieved through the following formula: ; in, σ H represents the trainable parameters of the high-frequency pathway; Step 2.2.3: Low frequency domain features L ( μ , v ) and high-frequency features H ( μ , v ) perform inverse Fourier transform respectively to obtain the low-frequency feature map f L ( x , y ) and high-frequency feature maps f H ( x , y ), which is specifically expressed by the following formula: ； ； Among them, FFT -1 represents the inverse Fourier transform; Then, the low-frequency feature map and the high-frequency feature map corresponding to the first feature map, the low-frequency feature map and the high-frequency feature map corresponding to the second feature map, the low-frequency feature map and the high-frequency feature map corresponding to the third feature map, and the low-frequency feature map and the high-frequency feature map corresponding to the fourth feature map are obtained.

4. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 2, characterized in that: Step 2.3 specifically includes: Step 2.3.1: Perform a high-frequency feature fusion operation on the fourth feature map and the high-frequency feature map corresponding to the fourth feature map to obtain a first high-frequency fused feature map; Step 2.3.2: Perform a bidirectional positioning operation on the edge information of the first high-frequency fusion feature map to obtain a first contour map; Step 2.3.3: Fuse the first contour map with the third feature map to obtain a first contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the high-frequency feature maps corresponding to the first contour fusion feature map and the third feature map to obtain a second contour map. Step 2.3.4: Fuse the second contour map with the second feature map to obtain a second contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the second contour fusion feature map and the high-frequency feature map corresponding to the second feature map to obtain a third contour map. Step 2.3.5: Fuse the third contour map with the first feature map to obtain a third contour fusion feature map. Similarly, perform a high-frequency feature fusion operation and an edge information bidirectional positioning operation on the third contour fusion feature map and the high-frequency feature map corresponding to the first feature map to obtain a fourth contour map. Step 2.3.6: Perform convolution on the fourth contour map to obtain the edge contour prediction map.

5. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 4, characterized in that: Step 2.3.1 specifically includes: The high-frequency feature map corresponding to the fourth feature map is subjected to 1x1 convolution to obtain the first convolution feature map, the first convolution feature map and the fourth feature map are spliced ​​in the channel dimension to obtain the first spliced ​​feature map, the first spliced ​​feature map is processed by the channel attention mechanism to obtain the first spatial weight map, the first spatial weight map is feature added with the fourth feature map to obtain the first high-frequency fusion feature map.

6. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 4, characterized in that: Step 2.3.2 specifically includes: Perform 1x1 convolution on the first high-frequency fusion feature map to obtain a second convolution feature map, process the second convolution feature map through the foreground module to obtain a target foreground feature map, specifically, process the second convolution feature map respectively through three 3x3 convolutions to obtain a first intermediate feature map, a second intermediate feature map and a third intermediate feature map, perform dilated convolution with a dilation rate of 1 on the first intermediate feature map to obtain a first foreground feature map, add features of the first foreground feature map and the second intermediate feature map to obtain a fourth intermediate feature map, perform dilated convolution with a dilation rate of 2 on the fourth intermediate feature map to obtain a second foreground feature map, add features of the second foreground feature map and the third intermediate feature map to obtain a fifth intermediate feature map, perform dilated convolution with a dilation rate of 3 on the fifth intermediate feature map to obtain a third foreground feature map, perform channel-level splicing on the first foreground feature map, perform 3x3 convolution and normalization on the spliced ​​feature map to obtain a target foreground feature map; The second convolution feature map is processed by a background module to obtain a target background feature map. The background module has the same operation as the foreground module, but the weight matrix of the dilated convolution and the convolution is different; The first contour map is obtained by element-wise multiplication of the target foreground feature map and the target background feature map.

7. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 2, characterized in that: Step 2.4 specifically includes: Step 2.4.1: Process the fourth feature map and the low-frequency feature map corresponding to the fourth feature map using information enhancement, low-frequency feature fusion, information filtering, and information localization to obtain a first localization feature map. The first localization feature map and the edge contour prediction map are concatenated in the channel dimension and subjected to channel attention, convolution, and upsampling to obtain a first camouflaged animal prediction map. Step 2.4.2: Process the third feature map and the low-frequency feature map corresponding to the third feature map using information enhancement, low-frequency feature fusion, information filtering, and information positioning to obtain a second positioning feature map. Use information correction to process the second positioning feature map and the first camouflaged animal prediction map to obtain a first information-corrected feature map. Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and perform channel attention, convolution, and upsampling to obtain a second camouflaged animal prediction map. Step 2.4.3: Fuse the first information correction feature map with the second feature map to obtain a first fused feature map. Based on information enhancement, low-frequency feature fusion operations, information filtering, and information positioning, process the low-frequency feature maps corresponding to the first fused feature map and the second feature map to obtain a third positioning feature map. Based on the second camouflaged animal prediction map, perform information correction on the third positioning feature map to obtain a second information correction feature map. Concatenate the second information correction feature map with the edge contour prediction map in the channel dimension and perform channel attention, convolution, and upsampling to obtain a third camouflaged animal prediction map. Step 2.4.4: Fuse the second information correction feature map with the first feature map to obtain a second fused feature map. Based on information enhancement, low-frequency feature fusion operation, information filtering and information positioning, process the second fused feature map and the low-frequency feature map corresponding to the first feature map to obtain a fourth positioning feature map. Based on the third camouflaged animal prediction map, perform information correction on the fourth positioning feature map to obtain a third information correction feature map. The third information correction feature map and the edge contour prediction map are spliced ​​in the channel dimension and subjected to channel attention, convolution and upsampling to obtain the fourth camouflaged animal prediction map.

8. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 7, characterized in that: Step 2.4.1 specifically includes: Step 2.4.1.1: Perform feature enhancement on the fourth feature map to obtain a first enhanced feature map; Specifically, 1 x 1 convolution, first dilated convolution, second dilated convolution, third dilated convolution and global average pooling are performed on the fourth feature map, and the obtained feature maps are spliced. 1 x 1 convolution is performed on the spliced ​​feature map to obtain a first enhanced feature map, wherein the expansion rate of the first dilated convolution, the expansion rate of the second dilated convolution and the expansion rate of the third dilated convolution are all different; Step 2.4.1.2: Perform a low-frequency feature fusion operation on the low-frequency feature maps corresponding to the first enhanced feature map and the fourth feature map to obtain a first low-frequency fused feature map; Specifically, a 1 x 1 convolution is performed on the low-frequency feature map corresponding to the fourth feature map to obtain a second convolution feature map, the second convolution feature map is spliced ​​with the first enhanced feature map in the channel dimension to obtain a second spliced ​​feature map, the second spliced ​​feature map is processed by the channel attention mechanism to obtain a second spatial weight map, and the second spatial weight map is feature added with the first enhanced feature map to obtain a first low-frequency fusion feature map; Step 2.4.1.3: Filter the first low-frequency fusion feature map to obtain a first filtered feature map; Specifically, the first low-frequency fusion feature map is subjected to 1x1 convolution, channel reorganization, and channel grouping to obtain m group features, the m group features are subjected to channel maximum pooling to obtain the maximum response of the intra-group feature, and the m group features are subjected to channel average pooling to obtain the average response of the intra-group feature. The maximum response of the intra-group feature and the average response of the intra-group feature are spliced ​​in the channel dimension and passed through 3x3 convolution to obtain a convolved feature map, and the convolved feature map is feature-added with the first low-frequency fusion feature map to obtain a first filtered feature map; Step 2.4.1.4: Perform information positioning on the first filtering feature map to obtain a first positioning feature map; Specifically, the first filtered feature map is subjected to 3x3 convolution, normalization and ReLU activation function to obtain the first precise positioning feature map; the first filtered feature map is subjected to normalization and multi-head conversion attention to obtain the global positioning feature map; The first filter feature map is split to obtain a first split feature map and a second split feature map, where the size of the first split feature map is different from the size of the second split feature map, and the first split feature map is subjected to 3x3 convolution, normalization and Relu activation function to obtain a first local feature map. At the same time, the first split feature map is subjected to 3x3 convolution, normalization and Relu activation function to obtain a second local feature map, where the parameters of the two 3x3 convolutions of the first split feature map are different, the first local feature map and the second local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a second precise positioning feature map; The second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a third local feature map. At the same time, the second split feature map is subjected to 5x5 convolution, normalization and Relu activation function to obtain a fourth local feature map, wherein the parameters of the two 5x5 convolutions of the second split feature map are different; the third local feature map and the fourth local feature map are spliced, and the spliced ​​feature map is subjected to 1x1 convolution to obtain a third precise positioning feature map; Add the first precise positioning feature map, the second precise positioning feature map, the third precise positioning feature map, and the global positioning feature map, and perform 3x3 convolution on the added feature map to obtain a global-local feature map. Add the global-local feature map and the first filtered feature map to obtain a first positioning feature map. Step 2.4.1.5: The first positioning feature map and the edge contour prediction map are spliced ​​in the channel dimension and subjected to channel attention, convolution and upsampling to obtain the first camouflaged animal prediction map.

9. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 7, characterized in that: Step 2.4.2 specifically includes: Step 2.4.2.1: Process the third feature map and the low-frequency feature map corresponding to the third feature map through information enhancement, low-frequency feature fusion operations, information filtering, and information positioning to obtain a second positioning feature map; Specifically, the third feature map is enhanced to obtain a second enhanced feature map, a low-frequency feature fusion operation is performed on the second enhanced feature map and the low-frequency feature map corresponding to the third feature map to obtain a second low-frequency fused feature map, information is filtered on the second low-frequency fused feature map to obtain a second filtered feature map, and information positioning is performed on the second filtered feature map to obtain a second positioned feature map; Step 2.4.2.2: Based on the first camouflaged animal prediction image, perform information correction and edge contour information fusion on the second positioning feature image to obtain a first information correction feature image; Specifically, based on the first camouflaged animal prediction map, the first value UR ( x,y ), which is specifically expressed by the following formula: ； in, is the two-dimensional coordinate in the first camouflaged animal prediction map ( x , y ) at the eigenvalue; Based on the first numerical value, the first camouflaged animal prediction map is divided into an identified area and an uncertain area, wherein the first numerical value corresponding to each coordinate in the identified area is 0, and the first numerical value corresponding to each coordinate in the uncertain area is 1; The second positioning feature map is element-wise multiplied with the uncertain area, and the multiplied feature map is input into the residual spatial attention to obtain an unidentified positioning feature map. The second positioning feature map is element-wise multiplied with the identified area, and the multiplied feature map is input into the residual channel attention to obtain an identified positioning feature map. The second positioning feature map is input into the residual spatial attention and the residual channel attention respectively to obtain a spatial attention feature map and a channel attention feature map. The unidentified positioning feature map, the identified positioning feature map, the spatial attention feature map and the channel attention feature map are spliced, and the spliced ​​result is subjected to a 1x1 convolution to obtain a first information correction feature map. Step 2.4.2.3: Concatenate the first information-corrected feature map and the edge contour prediction map in the channel dimension and obtain the second camouflaged animal prediction map through channel attention, convolution, and upsampling.

10. The method for detecting camouflaged animals based on frequency domain decoupling and edge information fusion according to claim 1, characterized in that: In step 3, the total loss function value is calculated based on the first camouflaged animal prediction map, the second camouflaged animal prediction map, the third camouflaged animal prediction map, the fourth camouflaged animal prediction map, the true annotation and the edge contour prediction map, which is specifically implemented by the following formula: ； in, L All Represents the total loss function value, Indicates the i Camouflaged animal prediction map, G Indicates the true markup, Indicates the i Segmentation loss between the camouflaged animal predictions and the ground-truth annotations, Indicates the i The area overlap loss between the camouflaged animal prediction map and the true annotation, Represents the edge contour prediction map and the true edge The edge alignment loss between To camouflage the outline of the animal in the target image; Among them, the parameters in the frequency domain decoupling and edge information fusion deep neural network FDEF-Net model are iteratively updated, including the parameters in the image feature extractor, the trainable parameters of the high-frequency path and the trainable parameters of the low-frequency path in the image frequency domain decoupler, the parameters of the convolution in the edge contour predictor, and the parameters of the convolution in the camouflaged animal detector and the parameters of the multi-head conversion attention.

Images