Camouflage target detection method and device based on cross-domain fusion

Abstract

The invention relates to a camouflage target detection method and device based on cross-domain fusion, electronic equipment and a storage medium. The method comprises the following steps: respectively performing spatial domain and frequency domain feature extraction processing on an input image to obtain a spatial domain feature group and a frequency domain feature group; performing spatial feature fusion processing on the spatial domain feature group to obtain spatial fusion features, and performing frequency domain feature fusion processing on the frequency domain feature group to obtain frequency domain fusion features; and detecting a camouflage target in the input image based on the spatial domain fusion feature and the frequency domain fusion feature. The target detection precision can be effectively improved.

Description

Technical Field

[0001] This disclosure relates to image processing technology, and in particular to a method and apparatus for detecting camouflaged targets, electronic equipment, and storage medium. Background Technology

[0002] Camouflaged object detection (COD) is an important research area in computer vision, aiming to automatically identify and locate target objects in complex natural scenes that are highly similar to the background in terms of color, texture, or structural features. These targets typically possess significant concealment and are widely used in applications such as natural ecological monitoring, military reconnaissance, medical image analysis, and industrial inspection, placing high demands on the feature representation and discrimination capabilities of detection algorithms. Most existing COD methods are based on deep learning frameworks, using structures such as convolutional neural networks or visual Transformers to extract multi-layer features from the input image and employing multi-scale feature fusion, contextual modeling, or attention mechanisms to enhance the perception of target regions. However, because camouflaged targets and backgrounds exhibit high similarity in the spatial domain, relying solely on spatial domain features is often insufficient to effectively distinguish between targets and backgrounds, easily leading to missed or false detections. Therefore, it is necessary to consider incorporating frequency domain information into the object detection field to optimize detection accuracy. Summary of the Invention

[0003] This disclosure proposes a method, apparatus, electronic device, and storage medium for camouflaged target detection based on cross-domain fusion, which addresses the problem of low detection rate of camouflaged targets in the prior art.

[0004] According to a first aspect of this disclosure, a method for detecting camouflaged targets based on cross-domain fusion is provided, comprising: performing spatial domain and frequency domain feature extraction processing on an input image respectively to obtain a spatial domain feature group and a frequency domain feature group; performing spatial feature fusion processing on the spatial domain feature group to obtain spatial fusion features, and performing frequency domain feature fusion processing on the frequency domain feature group to obtain frequency fusion features; and detecting camouflaged targets in the input image based on the spatial domain fusion features and the frequency domain fusion features.

[0005] In some possible implementations, the step of performing spatial domain and frequency domain feature extraction processing on the input image to obtain spatial domain feature sets and frequency domain feature sets includes: The input image is processed using a multi-scale feature extraction model to obtain the spatial domain feature set. The frequency domain feature set is obtained by performing frequency feature extraction processing on the input image, and the frequency domain feature set includes high-frequency features and low-frequency features.

[0006] In some possible implementations, performing spatial feature fusion processing on the spatial domain feature group to obtain spatial fused features includes: Perform feature alignment processing on each spatial feature within the spatial domain feature group to obtain aligned spatial features; Channel compression and weighted attention prediction are performed on the alignment spatial features respectively to obtain the weighted attention of the corresponding spatial features; The weighted attention and the alignment space features are multiplied to obtain the weighted space features; The spatial features are summed to obtain the spatial fusion features.

[0007] In some possible implementations, frequency domain feature fusion processing is performed on the frequency domain feature group to obtain frequency domain fused features, including: A first fusion process is performed on the high-frequency feature group in the frequency domain feature group to obtain a first fused feature, and a second fusion process is performed on the low-frequency feature group in the frequency domain feature group to obtain a second fused feature; Frequency domain fusion features are obtained based on the first fusion feature and the second fusion feature.

[0008] In some possible implementations, a second fusion process is performed on the low-frequency feature group to obtain a second fused feature, including: The low-frequency feature maps in the low-frequency feature group are subjected to connection processing in the height and width directions respectively to obtain the first low-frequency connection feature and the second low-frequency connection feature; Feature encoding processing is performed on the first low-frequency connection feature and the second low-frequency connection feature respectively to obtain the first encoded feature and the second encoded feature; The second fused feature is obtained by fusing the first encoding feature and the second encoding feature.

[0009] In some possible implementations, a first fusion process is performed on the high-frequency feature group to obtain a first fused feature, including: The high-frequency feature maps of the high-frequency feature group are subjected to connection fusion processing in the height direction, width direction, and diagonal direction, respectively, to obtain high-frequency connection features; Attention processing is performed on the high-frequency connectivity features to obtain high-frequency attention features; The first fused feature is obtained by performing feature transformation processing on the high-frequency attention feature.

[0010] In some possible implementations, detecting camouflaged targets in the input image based on the spatial domain fusion features and the frequency domain fusion features includes: Determine the detection weights for the spatial domain fusion features and the frequency domain fusion features; Based on the detection weights, a weighted sum of the spatial domain fusion features and the frequency domain fusion features is performed to obtain weighted fusion features; The location of the camouflaged target is determined using the weighted fusion features.

[0011] According to a second aspect of this disclosure, a camouflage target detection device is provided, comprising: The cross-domain feature extraction module is used to perform spatial domain and frequency domain feature extraction processing on the input image to obtain spatial domain feature groups and frequency domain feature groups. The fusion module is used to perform spatial feature fusion processing on the spatial domain feature group to obtain spatial fused features, and to perform frequency domain feature fusion processing on the frequency domain feature group to obtain frequency domain fused features; The detection module is used to detect camouflaged targets in the input image based on the spatial domain fusion features and the frequency domain fusion features.

[0012] According to a third aspect of this disclosure, an electronic device is characterized by comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method described in any one of the first aspects.

[0013] According to a fourth aspect of this disclosure, a computer-readable storage medium has computer program instructions stored thereon, characterized in that the computer program instructions, when executed by a processor, implement the method described in any one of the first aspects.

[0014] Based on the embodiments of this disclosure, the spatial domain fusion features and frequency domain fusion features of the input image are integrated into the same camouflage target detection framework for processing. By modeling the spatial domain fusion features, the edge details and local difference information of the camouflage target in the image can be effectively characterized; while by modeling the frequency domain fusion features, the overall structure and semantic information of the camouflage target can be stably depicted. On this basis, the spatial domain fusion features and frequency domain fusion features are further fused to achieve the synergistic expression of detail information and structural information, which is ultimately used for the detection of camouflage targets in the input image. In this way, the ability to perceive subtle differences between camouflage targets and complex backgrounds can be effectively enhanced, thereby improving the accuracy and robustness of camouflage target detection.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0016] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0018] Figure 1 This is a flowchart of the camouflage target detection method in the embodiments of this disclosure; Figure 2 This is a schematic diagram of the spatial domain feature fusion model structure in an embodiment of this disclosure; Figure 3 This is a schematic diagram of the channel compression and fusion model. Figure 4 This is a schematic diagram of the frequency domain feature fusion model structure in an embodiment of this disclosure. Detailed Implementation

[0019] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0020] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0021] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0022] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0023] The method disclosed herein can be executed by an image processing device, such as a terminal device, a server, or other processing device. The terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, the method can be implemented by a processor calling computer-readable instructions stored in memory.

[0024] It is understood that the various method embodiments mentioned above in this disclosure can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this disclosure will not elaborate further.

[0025] Figure 1 This is a flowchart of the camouflage target detection method in the embodiments of this disclosure; as follows: Figure 1 As shown, the camouflaged target detection method includes: S10: Perform spatial domain and frequency domain feature extraction processing on the input image to obtain spatial domain feature groups and frequency domain feature groups; In some possible implementations, a color image of the camouflaged object can be obtained using an image acquisition device, which is the image in the example of this disclosure. Specifically, the input image in the embodiment of this disclosure includes the camouflaged object to be detected by this disclosure. The camouflaged object can be any species, object, or medical detection target, for example, it can be a rabbit, but this is not a specific limitation of the embodiment of this disclosure.

[0026] In this embodiment, the input image is first processed by spatial domain and frequency domain feature extraction to obtain spatial domain feature sets and frequency domain feature sets. Specifically, spatial domain features are extracted using the PvT_v2 method, which effectively captures the spatial arrangement features of pixel values ​​in the image, including information such as the shape, texture, and edges of the target, thereby characterizing the spatial structure of the image. Frequency domain features are extracted using wavelet transform, which decomposes the image into different frequency components, making it particularly suitable for revealing high-frequency details and low-frequency structures in an image. Through the extraction of frequency domain features, the detailed parts of the image can be analyzed more comprehensively.

[0027] S20: Perform spatial feature fusion processing on the spatial domain feature group to obtain spatial fused features, and perform frequency domain feature fusion processing on the frequency domain feature group to obtain frequency domain fused features; In this embodiment, feature fusion processing can be performed on the spatial domain feature group and the frequency domain feature group respectively to obtain spatial fused features and frequency domain fused features. For the spatial domain feature group, a multi-scale adaptive feature fusion method is used to refine the features using convolutional kernels of different sizes, and then an appropriate scale is adaptively selected for fusion to obtain accurate spatial fused features. For the frequency domain feature group, it is decomposed into high-frequency and low-frequency feature groups using a frequency domain extraction method. Attention enhancement processing is applied to the high-frequency features to obtain high-frequency fused features; the low-frequency features are processed through concatenation and encoding to obtain low-frequency fused features. Finally, by combining the high-frequency and low-frequency fused features, a frequency domain fused feature is generated.

[0028] S30: Detect camouflaged targets in the input image based on the spatial domain fusion features and the frequency domain fusion features.

[0029] In this embodiment, camouflaged targets in the input image are detected based on the spatial domain fusion features and the frequency domain fusion features. Specifically, by fusing the spatial domain fusion features and the frequency domain fusion features, the advantages of both are combined to further improve the detection capability of camouflaged targets. The fused features are then fed into a subsequent detection module for accurate localization of camouflaged targets in the input image.

[0030] Based on the above configuration, this embodiment integrates the spatial domain fusion features and frequency domain fusion features of the input image into the same camouflage target detection framework for processing. Modeling the spatial domain fusion features effectively characterizes the edge details and local differences of camouflage targets in the image; while modeling the frequency domain fusion features stably depicts the overall structure and semantic information of the camouflage target. Furthermore, the spatial domain fusion features and frequency domain fusion features are fused to achieve a synergistic expression of detail and structural information, ultimately used for the detection of camouflage targets in the input image. This approach effectively enhances the perception of subtle differences between camouflage targets and complex backgrounds, thereby improving the accuracy and robustness of camouflage target detection.

[0031] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings. First, an input image can be obtained in the embodiments of this disclosure. The input image is a color image including a disguised target object. In the embodiments of this disclosure, the method of obtaining the input image may include at least one of the following methods: A) Directly use an image acquisition device to acquire color images including camouflaged target objects; wherein, the image acquisition device can be a color camera, video recording device or other device with image acquisition capabilities, and the specific type of image acquisition device is not limited in this disclosure embodiment.

[0032] B) Transmitting and receiving color images including camouflaged target objects via electronic devices; in this embodiment of the disclosure, color images transmitted by other electronic devices can be received via wired and / or wireless communication, and the communication method is not a specific limitation of this disclosure.

[0033] C) Reading a color image including a disguised target object from a storage medium or database; in this embodiment of the disclosure, a pre-stored color image can be read from a local storage device or server according to a received read instruction, and this disclosure does not specifically limit this.

[0034] It should be noted that the input images in the embodiments of this disclosure may come from the same scene or different scenes, and the type, shape and background environment of the disguised target object in the image can be set according to the actual application requirements. Those skilled in the art can select the corresponding image source according to the specific application scenario, and no specific limitation is made here.

[0035] After obtaining a color image including a disguised target object, disguised target object detection processing can be performed on the color image. Specifically, in this embodiment of the disclosure, the color image can be used as input and fed into a pre-built disguised target object detection model for processing, so as to realize the identification and location of the disguised target object in the image; The camouflaged target detection model in this embodiment is constructed based on spatial feature analysis and frequency feature analysis, including multiple feature processing paths for different features. This embodiment can effectively detect camouflaged targets by simultaneously utilizing detailed texture information and overall structural semantic information in an input image. Those skilled in the art can adjust the number, resolution, or source of the input image according to actual application needs, but this is not intended to limit the scope of this disclosure.

[0036] In some possible implementations, performing spatial domain and frequency domain feature extraction processing on the input image to obtain a spatial domain feature group and a frequency domain feature group includes: performing multi-scale feature extraction processing on the input image using a multi-scale feature extraction model to obtain the spatial domain feature group; and performing frequency feature extraction processing on the input image to obtain the frequency domain feature group, wherein the frequency domain feature group includes high-frequency features and low-frequency features.

[0037] For spatial domain feature extraction, embodiments of this disclosure utilize a multi-scale feature extraction method to perform multi-scale feature extraction processing on the input image, obtaining the spatial domain feature set. The spatial domain feature set includes image features at different scales, each with different spatial resolutions. For example, ResNet networks, pyramid feature extraction networks, and PVT network architectures can be used to extract multi-scale spatial features, resulting in the spatial domain feature set. Embodiments of this disclosure can employ a PVT-V2 network to perform multi-scale feature extraction, for example, obtaining four spatial features at different scales. Each scale's spatial features can focus on a suitable context, effectively solving the problem of feature fusion at different scales in camouflage scenes and ensuring that spatial domain features accurately represent important information in the image.

[0038] For frequency domain feature extraction, in some possible implementations, Fourier transform or wavelet transform can be performed on the input image to obtain the high-frequency feature group and the low-frequency feature group.

[0039] This disclosure allows for the performance of wavelet transform or Fourier transform on the input image to obtain high-frequency feature groups and low-frequency feature groups. The high-frequency feature groups contain detailed information about the image, while the low-frequency feature groups contain global information. It's important to note that, unlike the backbone network used in multi-scale feature extraction to obtain features at different scales, wavelet transform and Fourier transform achieve explicit frequency decoupling of features through predefined filters, clearly separating low-frequency structural components from high-frequency detailed components at the signal level. In contrast, the high and low-frequency characteristics corresponding to multi-scale features are implicit representation frequencies, derived from changes in spatial resolution and receptive field expansion, rather than strict signal frequency decomposition. The spatial and frequency domain feature methods employed in this disclosure enable feature fusion in both feature domains.

[0040] In this embodiment, the high-frequency feature group is used to characterize high-frequency information related to edge changes and texture details in the image, while the low-frequency feature group is used to characterize the overall structure, regional distribution, and global semantic information in the image. By separating the frequency features of the input image, different frequency components can be modeled and utilized separately in subsequent processing, thereby providing basic feature inputs for subsequent differentiated fusion processing of high-frequency and low-frequency features.

[0041] Given the spatial and frequency domain features, fusion processing can be performed on the features in each domain. The spatial feature fusion processing on the spatial domain feature group to obtain spatial fused features includes: performing feature alignment processing on each spatial feature within the spatial domain feature group to obtain aligned spatial features; performing channel compression and weighted attention prediction processing on the aligned spatial features to obtain weighted attention for the corresponding spatial features; performing a product processing on the weighted attention and the aligned spatial features to obtain weighted spatial features; and performing a summation processing on each weighted spatial feature to obtain the spatial fused features.

[0042] Figure 2 This is a block diagram of the spatial domain feature extraction model according to an embodiment of the present disclosure. The embodiments of the present disclosure first perform feature alignment processing on each spatial feature within the spatial domain feature group. This process ensures that all features can be processed and fused at the same spatial scale in subsequent feature fusion and weighting steps, avoiding potential mismatches between different scales. The alignment processing includes performing basic convolution processing and channel integration processing on the spatial features. The basic convolution processing includes convolution, batch normalization, and activation processing (such as ReLU). It is important to note that because the spatial features within the spatial feature group have different scales, the convolution parameters for each spatial feature in the basic convolution are different. For example, a 1×1 convolution kernel is used in the shallowest layer (first layer of spatial features), and larger-sized convolution kernels (3×3, 5×5, 7×7) are progressively used in deeper layers (second to fourth spatial features), thereby guiding low-level features to focus on local patterns, while high-level features integrate broader contextual information. This asymmetric design ensures that the effective receptive field of each layer matches its semantic role, rather than treating all layers as interchangeable. Channel integration processing employs 1×1 convolutions to project refined features into a unified embedding space. This configuration enables adaptive selection of the dominant fused representation scale at each spatial location.

[0043] Next, channel compression and weighted attention prediction are performed on the aligned spatial features. Figure 3 This is a schematic diagram of the channel compression and fusion structure. Channel compression is performed using 1×1 convolutions, which reduces the number of channels in the feature map and effectively extracts important information from each channel; meanwhile, the weighted attention prediction process dynamically predicts weights for each spatial location, reflecting the importance of each spatial location.

[0044] Specifically, for channel compression, since the original channel dimensions differ across layers, the refined features are projected onto a unified embedding space using a 1×1 convolution to obtain the channel-compressed features. For weighted attention prediction, a 1×1 convolution is performed on the channel-compressed features to obtain the weighted attention coefficients.

[0045] After aligning spatial features to a uniform spatial resolution, the most suitable scale is dynamically selected based on specific needs, thereby achieving location-adaptive feature fusion rather than averaging across all scales. For each scale layer, we parameterize the "1×1 compression" and "1×1 convolution" modules as two learnable mappings. Based on the aligned feature maps, the "compression" layer independently applies a 1×1 convolution (including BN and ReLU) to each spatial location.

[0046] In this process, the weighted attention prediction uses a 1×1 convolution to represent the weight strength that should be assigned to the feature scale at position (H,W). This weight is then broadcast and multiplied by the aligned spatial features to produce a scale-weighted response, resulting in weighted spatial features. Subsequently, the weighted spatial features are summed to obtain the spatial fusion features.

[0047] Using the above method, a set of weights spanning four scales is learned for each spatial location, so that the fused representation is dominated by the most informative scale, while naturally suppressing the less relevant responses.

[0048] In some optimized embodiments, to stabilize the fused representation and avoid introducing new noise, lightweight residual optimization is ultimately employed: two consecutive 3×3 convolutions with batch normalization are applied to the spatial fusion features, the output is added to the input, and then activated by ReLU. Finally, a 1×1 convolution is used to reduce the channel dimension, generating the optimized spatial fusion features. Based on this configuration, information from spatial domain features can be fully utilized in the camouflage target detection process, optimizing the fusion between features at different scales and improving overall detection performance, especially in the identification of complex backgrounds and subtle targets, where it has significant application value.

[0049] For frequency domain feature fusion, the frequency domain feature group is subjected to frequency domain feature fusion processing to obtain frequency domain fused features, including: performing a first fusion processing on the high-frequency feature group in the frequency domain feature group to obtain a first fused feature, and performing a second fusion processing on the low-frequency feature group in the frequency domain feature group to obtain a second fused feature; and obtaining frequency domain fused features based on the first fused feature and the second fused feature.

[0050] This embodiment of the invention performs a first fusion process on the high-frequency feature group, including connecting and fusing the high-frequency feature maps along the height, width, and diagonal directions, and then applying an attention mechanism to obtain high-frequency fused features. For the low-frequency feature group, a second fusion process is performed, first connecting the low-frequency feature maps along the height and width directions, and then encoding the connected features through a two-branch process to finally obtain low-frequency fused features. Then, based on the first and second fused features, a frequency domain fusion feature is obtained, which combines image details and global information. Finally, the frequency domain fusion feature is used to detect camouflaged target objects in the input image. This invention, through the fusion processing of multi-band features, can significantly improve the accuracy and robustness of camouflaged target object detection.

[0051] Performing a second fusion process on the low-frequency feature group to obtain a second fused feature includes: performing connection processing in the height and width directions on the low-frequency feature maps in the low-frequency feature group to obtain a first low-frequency connection feature and a second low-frequency connection feature; performing feature encoding processing on the first low-frequency connection feature and the second low-frequency connection feature to obtain a first encoded feature and a second encoded feature; and fusing the first encoded feature and the second encoded feature to obtain the second fused feature.

[0052] This embodiment of the disclosure can first perform connection processing on the low-frequency feature maps in the low-frequency feature group in both the height and width directions. Specifically, the low-frequency features It contains at least two low-frequency feature maps. To effectively fuse low-frequency feature information, these feature maps are concatenated. In the height direction, the rows of each low-frequency feature map are first concatenated to form the first low-frequency concatenated feature. In the width direction, the columns of each low-frequency feature map are further concatenated to obtain the second low-frequency concatenated feature. This connection processing allows the low-frequency feature map to be expanded in spatial dimensions, thereby preserving more image information and enhancing the effect of subsequent processing.

[0053] Subsequently, feature encoding processing can be performed on the first low-frequency connection feature and the second low-frequency connection feature respectively to obtain the first encoded feature and the second encoded feature. The feature encoding processing includes: performing a two-branch process on the input connection feature; wherein the first branch process includes: performing a linear transformation and activation processing on the connection feature; the second branch process includes: performing a linear transformation, feature dimensionality reduction, activation processing, and state space transformation processing on the connection feature; performing a product process and a linear transformation on the two features obtained from the two-branch process to obtain the encoded feature corresponding to the connection feature. The input connection feature is the first low-frequency connection feature and the second low-frequency connection feature. For ease of description, the following explanation uses the input connection feature as an example and will not elaborate further.

[0054] In this embodiment, after completing the connection processing of the low-frequency feature map, the obtained connection features are flattened to convert the feature map into one-dimensional data (B, H×W×C) for subsequent linear transformation and feature encoding. Next, feature encoding processing is performed on the first and second low-frequency connection features. Feature encoding includes performing a two-branch process on each connection feature. In the first branch, a linear transformation is first performed on the first low-frequency connection feature to obtain feature F, and then a nonlinear activation function is applied to this feature. For example, with Sigmoid, the activated feature A is the first encoded feature:

[0055] Where W and b are the weights and biases, respectively, and are learnable parameters.

[0056] Furthermore, in the second branch, the first low-frequency connection feature L is first linearly transformed to obtain the linearly transformed feature F. Then, a feature dimensionality reduction convolution operation is applied, using a convolution kernel size of 1×1, or alternatively 3×3, which is not specifically limited in this disclosure. Next, a nonlinear activation function is used to obtain nonlinear features, followed by a state space transformation to obtain the second encoded feature. The state space transformation process uses an SSM model, and the state equation and output equation used are set by those skilled in the art according to their needs, and are not specifically limited here. Through state space transformation, accurate representation of features in long-distance dimensions can be achieved, thereby improving model performance. Finally, the final state space transformed feature is obtained, which is the second encoded feature.

[0057] The linear transformations of the first and second branches use different weight parameters, which are learned during training. Further, a linear transformation is performed after multiplying the first and second encoded features to obtain the final first branch output features. Similarly, the second low-frequency connection features undergo the same encoding process to obtain the second branch output features. After feature encoding is complete, especially after multiple transformations (e.g., state-space transformation), the one-dimensional first branch output features and the first branch output features need to be restored to dimensional feature maps with the original input image, and then directly concatenated to obtain the final second fused feature.

[0058] This concatenation method preserves the independent information of each encoded feature and combines them to form a richer feature representation. In some optional implementations, the second fused feature can be further optimized, for example, by performing CBR (convolution, batch normalization, and ReLU activation) processing on the second fused feature. The CBR module extracts local features through convolution operations, improves training stability through batch normalization, and provides non-linear transformation through activation functions. To further enhance feature representation capabilities and avoid the gradient vanishing problem, residual connections are introduced into feature processing. In residual connections, low-frequency features are directly combined... The features are then added to the CBR-processed features to obtain the final optimized second fused features. These steps, through multi-level feature processing, nonlinear transformation, and information fusion, significantly improve the robustness and accuracy of camouflaged target detection.

[0059] Perform a first fusion process on the high-frequency feature group to obtain the first fused feature, including: The high-frequency feature maps of the high-frequency feature group are subjected to connection fusion processing in the height direction, width direction, and diagonal direction, respectively, to obtain high-frequency connection features; Attention processing is performed on the high-frequency connectivity features to obtain high-frequency attention features; The first fused feature is obtained by performing feature transformation processing on the high-frequency attention feature.

[0060] This embodiment of the disclosure targets a high-frequency feature group extracted through frequency domain transformation. First, it performs connection and fusion processing on multiple high-frequency feature maps contained within the high-frequency feature group in the height, width, and diagonal directions, respectively, thereby aggregating multi-directional texture detail information to obtain high-frequency connection features. Since high-frequency features contain more detailed relationships, this embodiment of the disclosure introduces more spatial connection methods in high-frequency feature fusion, further enriching the high-frequency feature information. Specifically, connection processing can be performed on the high-frequency feature maps of the high-frequency feature group in the height, width, and diagonal directions, respectively, to obtain a first high-frequency connection feature, a second high-frequency connection feature, and a third high-frequency connection feature; channel direction connection processing is then performed on the first, second, and third high-frequency connection features to obtain the high-frequency connection features.

[0061] In some possible implementations, after obtaining the high-frequency connection features, the embodiments of this disclosure can further perform fusion optimization on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features. The fusion optimization method includes any of the following: ① performing channel-direction connection processing on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features; ② calculating the distance between the first fusion feature and the high-frequency fusion connection features; if the distance is less than a distance threshold, using the first fusion feature as the optimized high-frequency feature; if the distance is greater than the distance threshold, performing channel-direction connection processing on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features. With the optimized high-frequency connection features obtained, the optimized high-frequency connection features are used as high-frequency connection features for subsequent attention processing.

[0062] Next, attention processing is performed on the high-frequency connectivity features to obtain high-frequency attention features. Specifically, firstly, recombination processing is performed on the high-frequency connectivity features to obtain recombination features; then, three-branch processing is performed on the recombination features to obtain first-branch features, second-branch features, and third-branch features; the product of the second-branch features and the third-branch features is used to obtain the first-branch product features; and the product of the first-branch features and the first-branch product features is used to obtain the high-frequency attention features.

[0063] The reorganization process for the high-frequency connectivity features includes: performing block processing on the high-frequency connectivity features in the width and height directions to obtain multiple block features; and performing connection processing on the block features in the height direction to obtain the reorganized features. The number of blocks can be set according to requirements, such as 2×2, or other numbers; this disclosure does not specifically limit this. This block processing strategy aims to achieve local attention enhancement: compared to global attention, which easily allows the camouflaged region to be submerged by background information, local attention allows the model to focus more on subtle differences within small regions, amplifying detailed textures and making it easier to detect edge changes of camouflaged objects; at the same time, block operations can reduce the computational complexity of the attention mechanism. Subsequently, the high-frequency connectivity features are input into the attention mechanism module for attention processing; during this process, the feature data undergoes dimensional reorganization, capturing short-distance dependencies and local statistical distribution characteristics between features, ultimately obtaining the first fusion feature for subsequent fusion calculations.

[0064] Upon obtaining the reconstructed features, multi-branch processing is performed on them. Within the attention processing unit, this embodiment improves the classic query, key, and value interaction logic by performing multi-level transformations on the features through statistical branches to enhance discriminability. Specifically, after performing reconstruction processing on the input feature stream, this embodiment generates Query, Key, and Value through three parallel branches. Specifically, Query is the first branch, Key is the second branch, and Value is the third branch.

[0065] The first and second branch processing methods respectively include feature centering, feature squarening, channel normalization, and height*width normalization. The purpose of feature centering is to highlight the "subtle differences" between the camouflage region and the background by eliminating the mean background. In this processing, the third branch is not centered to maintain the integrity of its original expression, while the second branch and its related paths are centered, which amplifies feature fluctuations and compresses flat background areas, thereby making subtle textures and edge changes clearer. In this embodiment, a feature square transformation is performed on the centered features. The square operation can achieve a contrast stretching effect of "larger values ​​are larger, smaller values ​​are smaller," so that the model pays more attention to texture boundaries, detail changes, and the outline of the camouflage, enhancing the response of significant features while suppressing weak background features, resulting in stronger recognition contrast. To further improve the robustness of the model in complex backgrounds, this embodiment performs double normalization processing on the channel dimension (C) and spatial dimension (H×W) after feature squarening. Among them, normalization along the channel dimension is used to constrain the activation amplitude of each semantic channel (such as texture, color, and edge channels), suppressing the contributions of different channels to the same order of magnitude, and preventing certain high-energy channels or noise channels from dominating the correlation calculation; normalization along the spatial dimension balances the energy distribution at each location in the image, making the contribution of each pixel location comparable, and effectively avoiding strong texture backgrounds or large areas from masking small camouflaged targets.

[0066] Furthermore, the product of the second branch features and the third branch features is used to obtain the first branch product features; the product of the first branch features and the first branch product features is used to obtain the high-frequency attention features. Given the high-frequency attention features, feature transformation processing is performed on them to obtain the first fused features. The feature transformation includes transforming the feature dimensions; through block processing in the width and height directions, the features are restored to the dimensions of the high-frequency features within the high-frequency feature group, thus obtaining the first fused features.

[0067] Based on the first fusion feature and the second fusion feature, a frequency domain fusion feature is obtained. Furthermore, given the first and second fusion features, the first and second fusion features are connected along the channel direction to obtain the frequency domain fusion feature. For example, the first and second fusion features can be added together to obtain the frequency domain fusion feature; or the first and second fusion features can be connected along the channel direction and then convolutional to obtain the frequency domain fusion feature.

[0068] The location of the disguised target object can then be obtained through spatial fusion features and frequency domain fusion features. The detection of the disguised target in the input image based on the spatial domain fusion features and the frequency domain fusion features includes: adding the spatial domain fusion features and the frequency domain fusion features to obtain a summed fusion feature, and using the summed fusion feature to detect the location of the target object. For example, activation processing can be directly performed on the summed fusion feature to obtain the location information of the disguised target object, where the location of the disguised target object is where the feature value is greater than 0.5 in the activated feature. Alternatively, convolution processing can be performed on the summed fusion feature followed by activation processing to obtain the location of the disguised target object. This disclosure does not specifically limit this approach.

[0069] In other embodiments, detecting a camouflaged target in the input image based on the spatial domain fusion feature and the frequency domain fusion feature includes: determining the detection weights of the spatial domain fusion feature and the frequency domain fusion feature; performing a weighted sum of the spatial domain fusion feature and the frequency domain fusion feature based on the detection weights to obtain a weighted fusion feature; and using the weighted fusion feature to determine the location of the camouflaged target.

[0070] In some possible implementations, embodiments of this disclosure first determine the detection weights of the spatial domain fusion features and the frequency domain fusion features. The allocation of these weights considers the importance of each feature in camouflage target detection and is dynamically adjusted according to feature differences in different scenarios to achieve optimal feature complementarity and information maximization in subsequent weighted fusion. Embodiments of this disclosure can pre-store the detection weights for each type of target object, allowing for the allocation of detection weights by reading the stored weight information. Furthermore, by weighting and fusing the spatial domain fusion features and the frequency domain fusion features, the advantages of both are combined, further enhancing the detection capability of camouflage targets. During the weighting and fusion process, spatial domain features and frequency domain features are effectively complementary, maximizing their information content. This weighting and fusion can be achieved through weighted summation or other fusion strategies, ensuring a balance and enhancement of local details in the spatial domain features and global structural information in the frequency domain features. The weighted fusion features are then fed into subsequent detection for accurate localization of camouflage targets in the input image. This weighted fusion process effectively suppresses background noise and highlights subtle differences between the camouflage target and the background, thereby significantly improving the accuracy and robustness of camouflage target detection. This method, through precise feature fusion and weight allocation, not only improves the ability to identify camouflaged targets, but also enhances its stability and adaptability in complex environments.

[0071] Compared with the prior art, the beneficial effects of this disclosure include the following aspects: 1. This invention achieves high-precision camouflage target capture based on frequency domain decoupling, overcoming the limitations of single spatial domain detection. The embodiments disclosed herein decompose the input image into high-frequency and low-frequency feature groups using wavelet transform, enabling parallel processing of the detailed texture and macroscopic semantics of the camouflage target from a frequency domain perspective. This method effectively solves the problem of blurred edges and difficulty in extraction caused by the high integration of color, texture, and background in the field of camouflage object detection, significantly improving the model's feature recognition capability in extreme camouflage environments.

[0072] 2. A State Space Transformation (SSM) architecture is introduced to achieve efficient modeling of long-distance semantic dependencies. In the low-frequency feature processing path, this embodiment employs an encoding module that includes state space transformation processing. Compared to traditional convolutional kernels, this architecture, through linear transformation, feature dimensionality reduction, and state space evolution, can acquire a wider receptive field and deeper semantic context information at a lower computational cost, effectively capturing the macroscopic logical relationship between the overall outline of the camouflaged target and the background environment, and improving the global consistency of the detection results.

[0073] 3. An improved attention mechanism based on local enhancement is proposed, significantly enhancing the sensitivity to detect subtle differences. Addressing the challenge of extremely small local differences in camouflaged targets, which are easily obscured by the background, this disclosure achieves precise magnification of subtle local textures by dividing high-frequency features into local regions and performing enhanced attention processing. Through improved Query-Key-Value interaction logic, combined with feature centering, square transformation, and dual normalization calibration of channels and space, the model can suppress strong background noise and highlight the weak edges of camouflaged areas. While maintaining the accuracy of dense prediction tasks, the computational complexity is significantly reduced through block-based computation.

[0074] 4. A multi-level, multi-dimensional feature fusion mechanism has been established to ensure the robustness and accuracy of the detection results. This disclosure employs a sophisticated fusion strategy to fuse attention-enhanced high-frequency detail features with low-frequency semantic features processed by state space transformation in multiple stages (such as first, second, and third fusion processes). This comprehensive processing flow, from local to global and from micro-texture to macro-semantics, enables the model to output accurate and complete camouflaged target detection results even in challenging scenarios such as varying target sizes, complex background textures, and extremely concealed camouflage areas.

[0075] Based on the embodiments of this disclosure, feature information is fully quantified and effectively fused in both the spatial and frequency domains, thereby improving the accuracy of target detection.

[0076] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0077] In addition, this disclosure also provides a disguised object detection device, electronic device, computer-readable storage medium, and program, all of which can be used to implement any of the disguised object detection methods provided in this disclosure. The corresponding technical solutions and descriptions are described in the corresponding records in the method section and will not be repeated here.

[0078] The disguised object detection device includes: a cross-domain feature extraction module, used to perform spatial domain and frequency domain feature extraction processing on the input image respectively to obtain a spatial domain feature group and a frequency domain feature group; a fusion module, used to perform spatial feature fusion processing on the spatial domain feature group to obtain spatial fusion features, and perform frequency domain feature fusion processing on the frequency domain feature group to obtain frequency domain fusion features; and a detection module, used to detect disguised targets in the input image based on the spatial domain fusion features and the frequency domain fusion features.

[0079] In some embodiments, the apparatus provided in this disclosure may have functions or include modules that can be used to perform the methods described in the above method embodiments. Specific implementations can be referred to the descriptions in the above method embodiments, and for brevity, will not be repeated here. This disclosure also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above methods. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0080] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A camouflage target detection method based on cross-domain fusion, characterized in that, The method comprises the following steps: performing spatial domain and frequency domain feature extraction processing on the input image respectively to obtain a spatial domain feature group and a frequency domain feature group; performing spatial feature fusion processing on the spatial domain feature group to obtain a spatial fusion feature, and performing frequency domain feature fusion processing on the frequency domain feature group to obtain a frequency domain fusion feature; detecting a camouflaged target in the input image based on the spatial domain fusion feature and the frequency domain fusion feature.

2. The method of claim 1, wherein, The method comprises the following steps: performing multi-scale feature extraction processing on the input image by using a multi-scale feature extraction model to obtain the spatial domain feature group; performing frequency domain feature extraction processing on the input image to obtain the frequency domain feature group, wherein the frequency domain feature group comprises high-frequency features and low-frequency features.

3. The method of claim 1, wherein, The method comprises the following steps: performing feature alignment processing on each spatial feature in the spatial domain feature group to obtain aligned spatial features; performing channel compression and weight attention prediction processing on the aligned spatial features respectively to obtain weight attention of the corresponding spatial features; performing product processing on the weight attention and the aligned spatial features to obtain weighted spatial features; performing summation processing on each weighted spatial feature to obtain the spatial fusion feature.

4. The method of claim 1, wherein, The method comprises the following steps: performing first fusion processing on the high-frequency feature group in the frequency domain feature group to obtain a first fusion feature, and performing second fusion processing on the low-frequency feature group in the frequency domain feature group to obtain a second fusion feature; obtaining the frequency domain fusion feature based on the first fusion feature and the second fusion feature.

5. The method of claim 4, wherein, The method comprises the following steps: performing connection processing on the low-frequency features in the low-frequency feature group in the height and width directions respectively to obtain first low-frequency connection features and second low-frequency connection features; performing feature encoding processing on the first low-frequency connection features and the second low-frequency connection features respectively to obtain first encoded features and second encoded features; fusing the first encoded features and the second encoded features to obtain the second fusion feature.

6. The method according to claim 4 or 5, characterized in that, The method comprises the following steps: performing connection fusion processing on the high-frequency feature maps of the high-frequency feature group in the height direction, the width direction, and the diagonal direction to obtain high-frequency connection features; performing attention processing on the high-frequency connection features to obtain high-frequency attention features; performing feature conversion processing on the high-frequency attention features to obtain the first fusion feature.

7. The method of claim 1, wherein, The method comprises the following steps: determining detection weights of the spatial domain fusion feature and the frequency domain fusion feature; performing weighted summation processing on the spatial domain fusion feature and the frequency domain fusion feature based on the detection weights to obtain a weighted fusion feature; determining the position of the camouflaged target by using the weighted fusion feature.

8. A camouflage target detection apparatus characterized by comprising: The method comprises the following steps: The cross-domain feature extraction module is configured to perform spatial domain and frequency domain feature extraction processing on the input image respectively to obtain a spatial domain feature group and a frequency domain feature group; The fusion module is configured to perform spatial feature fusion processing on the spatial domain feature group to obtain spatial fusion features, and perform frequency domain feature fusion processing on the frequency domain feature group to obtain frequency domain fusion features; The detection module is configured to detect a camouflaged target in the input image based on the spatial domain fusion features and the frequency domain fusion features.

9. An electronic device, comprising: The computer program product comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images