A camouflage target detection method and system based on a dual-domain fusion enhanced network

Abstract

The application discloses a camouflage target detection method and system based on a dual-domain fusion enhanced network, and relates to the technical field of target detection.The application captures key difference features of a frequency domain amplitude spectrum and a phase spectrum by combining spatial domain and frequency domain nonlinear mapping through a nonlinear dual-domain fusion module, solves the problem of performance decline in low-contrast scenes and the like in dependence on an RGB spatial domain, and improves target discrimination;relying on a lightweight scale perception modulation transformer and a dual-feature fusion module, multi-scale features are extracted and aligned and fused, semantic relationships are integrated by cross attention, and the problems of detail loss and boundary blur caused by scale diversity are overcome;an upper and lower context feature enhancement module incorporates cross attention and edge auxiliary injection, accumulates multi-layer feature integration, takes into account global boundaries and local structures, effectively reduces false positives and false negatives and edge roughness, and multi-layer auxiliary supervision further strengthens robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a camouflage target detection method and system based on a dual-domain fusion enhancement network. BACKGROUND

[0002] In the field of computer vision, camouflage target detection aims to accurately identify and segment the disguised target from a complex scene with a high degree of similarity to the background. The breakthrough of this technology is of great significance to key application fields such as ecological environment protection, agricultural pest monitoring, industrial defect detection, and medical image analysis. At the same time, with the continuous progress of technology and the rapid development of society, the demand for antagonism of camouflage and anti-camouflage technology in the field of security defense is increasing. Camouflaged objects disguise real targets as other forms or hide in the environment to evade detection by monitoring and identification systems, thereby posing higher challenges to intelligent visual systems.

[0003] Although in recent years the deep learning technology has driven significant progress in the field of camouflage target detection, existing camouflage target detection methods still have the following shortcomings:

[0004] Firstly, most methods excessively rely on RGB spatial domain features (such as texture contrast, edge response, etc.), however, in camouflage scenes, the target and the background have very low distinguishability in the spatial domain. Especially under the condition of low contrast or complex texture interference, the model is difficult to mine key difference features that can effectively distinguish foreground and background from local texture, resulting in a significant decrease in detection performance.

[0005] Secondly, camouflage targets generally have scale diversity. Existing methods mostly use simple feature concatenation or fixed weight fusion to integrate multi-scale information, ignoring the semantic complementary relationship between different scale features. Therefore, when dealing with complex scenes such as slender targets, dramatic scale changes, or multi-target occlusion, the model is prone to problems such as loss of detail information and blurred boundaries.

[0006] Finally, the boundaries of camouflage targets are usually fuzzy and discontinuous (such as transparent targets or partial occlusion), and existing decoders mostly rely on a single attention mechanism for feature correction, making it difficult to balance global boundary integrity and accurate segmentation of local fine structures. This results in false positives, false negatives, and rough edges in the prediction results, thereby limiting the overall precision and robustness of the model. SUMMARY

[0007] The present application aims to provide a camouflage target detection method and system based on a dual-domain fusion enhancement network to improve the technical problem of existing models that excessively rely on RGB spatial domain features and ignore the semantic complementary relationship between different scale features.

[0008] To achieve the above object of the application, the embodiments of the present application provide the following technical solutions:

[0009] A camouflage target detection method based on a dual-domain fusion enhanced network comprises:

[0010] An original detection scene image in which a camouflage target is located is acquired and preprocessed to obtain a detection scene image;

[0011] A lightweight scale perception modulation transformer is constructed based on a scale perception modulation multi-head attention mechanism and a multi-head mixed convolution, combined with a scale perception aggregation module;

[0012] Based on Fourier transform, nonlinear convolution and channel attention mechanism, a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit are generated to construct a nonlinear dual-domain fusion module;

[0013] A gating edge module and a spatial mixing unit are constructed; and a context feature enhancement module is constructed based on a hybrid attention mechanism and the spatial mixing unit;

[0014] A camouflage target detection model is constructed based on the lightweight scale perception modulation transformer, the nonlinear dual-domain fusion module, the gating edge module and the context feature enhancement module;

[0015] The detection scene image is processed by the camouflage target detection model to obtain a camouflage target detection result.

[0016] In the above scheme, the nonlinear dual-domain fusion module is used to capture the key difference features of the frequency domain amplitude spectrum and the phase spectrum by combining the spatial domain and the frequency domain nonlinear mapping, solve the problem of performance decline in low contrast scenes and the like depending on the RGB spatial domain, and improve the target discrimination; the lightweight scale perception modulation transformer and the dual-feature fusion module are used to extract and align multi-scale features, integrate semantic relationships by cross attention, and overcome the problems of detail loss and boundary blur caused by scale diversity; the context feature enhancement module integrates cross attention and edge auxiliary injection, and takes into account global boundaries and local structures, effectively reduces false positives and misses and edge roughness, and multi-layer auxiliary supervision further strengthens robustness.

[0017] Further, the lightweight scale perception modulation transformer comprises an initial feature extraction layer, a first scale perception modulation module, a first patch embedding layer, a second scale perception modulation module, a second patch embedding layer, a scale perception modulation multi-head attention module, a third patch embedding layer and a multi-head attention layer connected in sequence;

[0018] The first scale perception modulation module and the second scale perception modulation module have the same structure and each comprises a multi-head mixed convolution layer and a scale perception aggregation layer; the scale perception modulation multi-head attention module comprises scale perception modulation modules and multi-head attention blocks stacked alternately.

[0019] The first patch embedding layer and the second patch embedding layer have the same structure.

[0020] Further, the nonlinear dual-domain fusion module comprises a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit; the frequency domain nonlinear mapping unit comprises a Fourier transform layer, a first nonlinear convolution layer, a channel attention layer and an inverse Fourier transform layer; the spatial domain nonlinear mapping unit comprises a second nonlinear convolution layer, a transposed convolution layer and a first convolution layer.

[0021] Further, the context feature enhancement module comprises a hybrid attention layer, a spatial mixing unit, a second global average pooling layer, a fifth nonlinear convolution layer and a sixth nonlinear convolution layer; the spatial mixing unit comprises a first point-wise convolution layer, a fourth nonlinear convolution layer and a second point-wise convolution layer.

[0022] In the above scheme, the lightweight scale-aware modulation transformer can efficiently extract multi-scale features from shallow to deep layers through the series connection of the scale-aware modulation module and the patch embedding layer, solve the problem of scale diversity of camouflage targets, and guarantee the hierarchy of features and the adaptability of subsequent modules; the nonlinear dual-domain fusion module breaks through the limitation of a single RGB spatial domain by means of spatial domain (including nonlinear, transposed convolution) and frequency domain (including Fourier transform, channel attention) mapping units, accurately captures dual-domain difference features in low-contrast and complex texture scenes, improves the discrimination of targets and backgrounds, integrates global semantics and local details, and relieves boundary blur and detail loss.

[0023] Further, the method for constructing the camouflage target detection model comprises:

[0024] Based on the gating edge module, the nonlinear convolution layer and the context feature enhancement module, a first branch is constructed;

[0025] Based on the context feature enhancement module and the nonlinear convolution layer, a second branch, a third branch and a fourth branch are respectively constructed;

[0026] Based on the nonlinear dual-domain fusion module, the high-pass filter, the low-pass filter and the nonlinear multi-layer perception, a first detection sub-model is constructed;

[0027] Based on the lightweight scale-aware modulation transformer and the first branch to the fourth branch, a second detection sub-model is constructed;

[0028] Based on the first detection sub-model and the second detection sub-model, a camouflage target detection model is constructed.

[0029] In the above scheme, in the multi-branch design, the first branch introduces the gated edge module to strengthen the edge information of the camouflage target, and the second to fourth branches realize fine processing of features by relying on the context feature enhancement module and the nonlinear convolution layer. Multi-branch parallel adaptation adapts different scale features from shallow to deep layers, avoiding information loss of single path processing. The first detection sub-model breaks the dependence on a single RGB spatial domain by using a nonlinear dual-domain fusion module to capture dual-domain difference features, and combines a high-pass / low-pass filter to accurately separate details and structures, and a nonlinear multilayer perceptron to strengthen feature expression, solving the problem of performance decline and detail loss in low-contrast scenes. The second detection sub-model provides multi-scale basic features from shallow to deep layers by using a lightweight scale perception modulation transformer, and combines multi-branch to realize hierarchical refinement of features. Finally, the dual detection sub-model fusion allows dual-domain enhanced features and multi-branch refined features to work together to comprehensively improve the discrimination, detail retention capability and robustness of camouflage target detection, effectively reducing false positives, false negatives and edge roughness.

[0030] Further, the training process of the camouflage target detection model is:

[0031] Collect original training images of the detection scene and perform data enhancement to generate detection scene training images;

[0032] Input the detection scene training images into the lightweight scale perception modulation transformer to generate shallow multi-scale detail enhancement features, middle multi-scale semantic detail fusion features, middle scale perception global semantic association features, and deep global semantic aggregation features;

[0033] Input the shallow multi-scale detail enhancement features into the seventh nonlinear convolution layer of the first branch to generate first convolution features; input the middle multi-scale semantic detail fusion features into the seventh nonlinear convolution layer of the second branch to generate second convolution features; input the middle scale perception global semantic association features into the seventh nonlinear convolution layer of the third branch to generate third convolution features; and input the deep global semantic aggregation features into the seventh nonlinear convolution layer of the fourth branch to generate fourth convolution features;

[0034] Input the second convolution features to the fourth convolution features into the first detection sub-model to generate high-frequency features and dual-domain fusion features;

[0035] Input the first convolution features and high-frequency features into the gated edge module to generate edge feature maps;

[0036] input the first convolutional feature, the edge feature map and the dual-domain fusion feature into a context feature enhancement module of the first branch to generate a first detail enhancement feature; input the second convolutional feature, the edge feature map and the dual-domain fusion feature into a context feature enhancement module of the second branch to generate a second detail enhancement feature; input the third convolutional feature, the edge feature map and the dual-domain fusion feature into a context feature enhancement module of the third branch to generate a third detail enhancement feature; and input the fourth convolutional feature, the edge feature map and the dual-domain fusion feature into a context feature enhancement module of the fourth branch to generate a fourth detail enhancement feature;

[0037] input the first detail enhancement feature into an eighth nonlinear convolutional layer and a second convolutional layer of the first branch to generate a shallow layer fusion feature; input the second detail enhancement feature into an eighth nonlinear convolutional layer and a second convolutional layer of the second branch to generate a middle layer semantic detail fusion feature; input the third detail enhancement feature into an eighth nonlinear convolutional layer and a second convolutional layer of the third branch to generate a middle layer global semantic fusion feature; and input the fourth detail enhancement feature into an eighth nonlinear convolutional layer and a second convolutional layer of the fourth branch to generate a deep layer global fusion feature;

[0038] input the shallow layer fusion feature, the middle layer semantic detail fusion feature, the middle layer global semantic fusion feature and the deep layer global fusion feature into a detection head to obtain a camouflage target detection training result;

[0039] based on the camouflage target detection training result, adjust network parameters of the camouflage target detection model through deep supervision learning.

[0040] In the above scheme, the multi-scale features generated by the lightweight scale perception modulation transformer provide a hierarchical basis from details to semantics for subsequent multi-branch processing, adapting to the characteristics of various scales of camouflage targets. The high-frequency features generated by the first detection sub-model break through the limitations of a single spatial domain and solve the problem of performance degradation in low-contrast scene detection. The edge feature map generated by the gating edge module injects precise edge information for subsequent feature enhancement, alleviating the blurring of the boundaries of camouflage targets. The context feature enhancement module combines the edge map and the dual-domain fusion feature to refine the features of each branch, achieving the cooperation of global semantics and local details and reducing the loss of details. Finally, through deep supervision learning based on the detection result, the parameters are adjusted to optimize the model accurately, significantly improve the detection accuracy, effectively reduce false positives, missed detection and rough edges, and strengthen the robustness of the model.

[0041] Further, the generation of the high-frequency feature and the dual-domain fusion feature comprises:

[0042] input the second convolutional feature to the fourth convolutional feature into a nonlinear dual-domain fusion module to generate corresponding time-frequency global sensitive features;

[0043] The high-frequency features and the low-frequency features are input into a double-feature fusion layer to generate double-domain fusion features.

[0044] The high-frequency features and the low-frequency features are input into a double-feature fusion layer to generate double-domain fusion features.

[0045] In the above scheme, the second to fourth convolutional features are synchronously processed in the RGB spatial domain and the frequency domain (amplitude spectrum, phase spectrum) by the nonlinear double-domain fusion module to generate time-frequency global features sensitive to low-contrast and complex texture scenes. Then, high-frequency details (edges, textures) and low-frequency structures (overall semantics) are separated by high-pass / low-pass filtering, and the two are fused by the cross-attention mechanism, with the low-frequency providing positioning anchors for the high-frequency, and the high-frequency supplementing fine information for the low-frequency. The finally output double-domain fusion features have both local sharpness and global integrity, significantly improving the boundary clarity and multi-scale scene adaptability of the camouflage target detection.

[0046] Further, the processing process of the lightweight scale perception modulation transformer is:

[0047] The detection scene training image is input into an initial feature extraction layer to generate an initial feature; the initial feature is input into a first scale perception modulation module to generate a shallow multi-scale detail enhancement feature; the shallow multi-scale detail enhancement feature is input into a first patch embedding layer to generate a first sequence feature; the first sequence feature is input into a second scale perception modulation module to generate a middle multi-scale semantic detail fusion feature; the middle multi-scale semantic detail fusion feature is input into a second patch embedding layer to generate a second sequence feature; the second sequence feature is input into a scale perception modulation multi-head attention module to generate a middle scale perception global semantic association feature; the middle scale perception global semantic association feature is input into a third patch embedding layer to generate a third sequence feature; and the third sequence feature is input into a multi-head attention layer to generate a deep global semantic aggregation feature.

[0048] In the above scheme, the lightweight scale perception modulation transformer generates progressive features from shallow details to deep semantics through hierarchical processing: the scale perception modulation module captures multi-scale information to adapt to the scale diversity of the camouflage target; the patch embedding layer converts the features into serialized vectors to adapt to the subsequent attention mechanism; and the alternately stacked modules realize smooth transition from local to global features. Finally, high-quality basic features are provided for subsequent processing, which solves the problems of detail loss and global semantic ambiguity while ensuring efficiency, and improves the detection effect.

[0049] A camouflage target detection system based on a double-domain fusion enhancement network includes:

[0050] The acquisition module is used for acquiring an original detection scene image of a camouflage target and pre-processing to obtain a detection scene image; a first construction module is used for constructing a lightweight scale perception modulation transformer based on a scale perception modulation multi-head attention mechanism and a multi-head mixed convolution, in combination with a scale perception aggregation module; a second construction module is used for generating a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit based on a Fourier transform, a nonlinear convolution and a channel attention mechanism, and constructing a nonlinear dual-domain fusion module; a third construction module is used for constructing a gated edge module and a spatial mixing unit; a context feature enhancement module is constructed based on the mixed attention mechanism and the spatial mixing unit; a model construction module is used for constructing a camouflage target detection model based on the lightweight scale perception modulation transformer, the nonlinear dual-domain fusion module, the gated edge module and the context feature enhancement module; and a camouflage detection module is used for processing the detection scene image through the camouflage target detection model to obtain a camouflage target detection result.

[0051] In the above scheme, the system can efficiently detect the camouflage target through the cooperation of multiple modules: the system can adapt to the diversity of target scales, break through the limitation of a single spatial domain, strengthen the boundary information and the global-local correlation, reduce false positives and false negatives, and improve the detection accuracy, detail retention capability and adaptability in complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0053] Figure 1 The method flowchart in the embodiments of the present application;

[0054] Figure 2 The nonlinear dual-domain fusion module structure diagram in the embodiments of the present application;

[0055] Figure 3 The frequency domain nonlinear mapping unit structure diagram in the embodiments of the present application;

[0056] Figure 4 The spatial domain nonlinear mapping unit structure diagram in the embodiments of the present application;

[0057] Figure 5 The gated edge module structure diagram in the embodiments of the present application;

[0058] Figure 6 The context feature enhancement module structure diagram in the embodiments of the present application;

[0059] Figure 7 A second detection sub-model structure diagram in an embodiment of the present application is shown in FIG. 2.

[0060] Figure 8 A first detection sub-model structure diagram in an embodiment of the present application is shown in FIG. 1.

[0061] Figure 9 A double-feature fusion layer structure diagram in an embodiment of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0063] Referring to Figure 1 The embodiment provides a camouflage target detection method based on a double-domain fusion enhanced network. Figure 1 The execution subject of the method shown can be a software and / or a hardware device. The execution subject of the present application can include but is not limited to at least one of the following: a user equipment, a network equipment, and the like. The user equipment can include but is not limited to a computer, a smart phone, a personal digital assistant (PDA), and the above-mentioned electronic devices, and the like. The network equipment can include but is not limited to a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of computers or network servers based on cloud computing. The cloud computing is a kind of distributed computing, which is a super virtual computer composed of a loose coupled computer group. The embodiment is not limited in this regard.

[0064] As Figure 1 shown, a camouflage target detection method based on a double-domain fusion enhanced network includes:

[0065] S1, obtaining an original detection scene image in which a camouflage target is located and performing preprocessing to obtain a detection scene image;

[0066] Specifically, in order to ensure the consistency of the input size, all input images are subjected to normalization processing and are randomly cropped into image blocks with a size of 384x384.

[0067] S2, based on scale-aware modulation multi-head attention mechanism and multi-head mixed convolution, combined with scale-aware aggregation module, a lightweight scale-aware modulation transformer is constructed;

[0068] Specifically, the lightweight scale-aware modulation transformer (SMT-T) includes an initial feature extraction layer, a first scale-aware modulation module, a first patch embedding layer, a second scale-aware modulation module, a second patch embedding layer, a scale-aware modulation multi-head attention module, a third patch embedding layer, and a multi-head attention layer connected in sequence.

[0069] The first scale-aware modulation module and the second scale-aware modulation module have the same structure, both including a multi-head mixed convolution layer (Multi-Head Mixed Convolution, MHMC layer) and a scale-aware aggregation layer (Scale-Aware Aggregation, SAA layer). The MHMC layer divides the channels into multiple heads, and each head uses a deep separable convolution (such as 3x3, 5x5, etc.) of different sizes to capture multi-scale local features. The SAA layer integrates the outputs through channel shuffling and grouping fusion to achieve efficient interaction. After multi-scale feature extraction and aggregation, the first scale-aware modulation module generates a modulation matrix M, which is then multiplied element by element with the value feature V to achieve scale-aware modulation. The scale-aware aggregation module is the scale-aware aggregation layer.

[0070] Due to the inefficiency of the multi-head attention mechanism (MHA) in the shallow stage, the MHMC layer is designed by replacing the attention calculation with convolution to achieve linear complexity and emphasizing multi-scale feature capture to support spatial and channel modulation. Compared with MHA, the MHMC layer has fewer parameters and faster calculation, which can improve the adaptability of shallow feature extraction and seamlessly integrate with subsequent attention mechanisms to improve the overall performance of the model.

[0071] The scale-aware modulation multi-head attention module is the core structure adopted in the third stage of the lightweight scale-aware modulation transformer, which is not a single module but an innovative stacking strategy aimed at achieving a smooth transition from local feature modeling to global dependency capture. Its core structure is an alternating stack of a scale-aware modulation module (SAM) and a multi-head self-attention (MSA) block, forming a sequence of "SAM-MSA-SAM-MSA-...". This alternating stacking structure enables the lightweight SAM block, which is good at extracting multi-scale local context, to provide more information-rich features for the powerful MSA block that follows it, which is good at establishing long-range dependencies. The two work together to efficiently simulate the progressive evolution of the network receptive field within a single stage, ultimately optimizing computational efficiency while ensuring performance. The scale-aware modulation module (SAM) has the same structure as the first scale-aware modulation module.

[0072] S3, based on Fourier transform, nonlinear convolution and channel attention mechanism, generate a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit, and construct a nonlinear dual-domain fusion module;

[0073] Specifically, in the existing camouflage target detection method, there is a technical problem of difficulty in mining key difference features. The traditional method often only extracts features in the spatial domain, which leads to insufficient sensitivity to subtle differences (such as texture, color gradient, etc.) between the camouflage object and the background, and cannot effectively capture the key information (such as high-frequency edges or low-frequency overall structural differences) hidden in the frequency domain. To solve this problem, the present application introduces a nonlinear dual-domain fusion module. Because the difference features of the camouflage target often exist in the spatial domain and the frequency domain: the spatial domain is good at capturing local pixel relationships, while the frequency domain can separate the amplitude spectrum (representing energy distribution) and the phase spectrum (representing structural information) through Fourier transform, thereby more comprehensively mining the differences. For this purpose, the existing fusion mechanism is improved, and a structure for nonlinear enhancement in the spatial domain and the frequency domain is designed to generate a nonlinear dual-domain fusion module.

[0074] In addition, the existing method also has the technical problems of loss of detail information and blurring of boundaries. In the multi-scale extraction process of the traditional backbone network, down-sampling often leads to loss of high-frequency details (such as edges and textures), and the edge module is usually designed independently and cannot dynamically interact with the main features, resulting in insufficient utilization of edge information. To solve this problem, the present application selects high-frequency and low-frequency components and introduces an edge map auxiliary mechanism, which is based on the frequency domain separation principle of high-pass and low-pass filtering: the high-frequency component highlights the detail differences, and the low-frequency component retains the overall structure, and through separation, information flooding during mixed processing can be avoided; at the same time, considering that the edge is the key area of detail loss, an edge map is generated from the low-level features to achieve early injection and dynamically weighted high-frequency features.

[0075] Thus, as shown in Figure 2 , the nonlinear dual-domain fusion module includes a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit. As shown in Figure 3 , the frequency domain nonlinear mapping unit includes a Fourier transform layer, a first nonlinear convolution layer, a channel attention layer, and an inverse Fourier transform layer. As shown in Figure 4 , the spatial domain nonlinear mapping unit includes a second nonlinear convolution layer, a transposed convolution layer, and a first convolution layer.

[0076] S4, construct a gated edge module and a spatial mixing unit; based on a hybrid attention mechanism and the spatial mixing unit, construct a context feature enhancement module;

[0077] Specifically, the existing method faces the technical problem of difficulty in identifying the segmentation of global boundaries and local structures. Traditional decoders often only rely on upsampling and simple cascading, resulting in difficulty in accurately distinguishing global boundaries (such as the overall outline of an object) from local structures (such as internal texture segmentation), causing inaccurate segmentation. The present application constructs a context feature enhancement module to solve the technical problem of inaccurate image segmentation caused by the difficulty in accurately distinguishing global boundaries (such as the overall outline of an object) from local structures (such as internal texture segmentation).

[0078] As shown in Figure 5 , the gating edge module includes a first global average pooling layer, a fully connected layer, and a third nonlinear convolution layer.

[0079] As shown in Figure 6 , each context feature enhancement module includes a hybrid attention layer, a spatial mixing unit, a second global average pooling layer, a fifth nonlinear convolution layer, and a sixth nonlinear convolution layer. The spatial mixing unit includes a first pointwise convolution layer, a fourth nonlinear convolution layer, and a second pointwise convolution layer.

[0080] S5, based on the lightweight scale-aware modulation transformer, the nonlinear dual-domain fusion module, the gating edge module, and the context feature enhancement module, a camouflage target detection model is constructed;

[0081] The S5 includes:

[0082] S5-1, based on the gating edge module, the nonlinear convolution layer, and the context feature enhancement module, a first branch is constructed;

[0083] S5-2, based on the context feature enhancement module and the nonlinear convolution layer, a second branch, a third branch, and a fourth branch are respectively constructed;

[0084] S5-3, based on the nonlinear dual-domain fusion module, high-pass filtering, low-pass filtering, and a nonlinear multilayer perceptron, a first detection submodel is constructed;

[0085] S5-4, based on the lightweight scale-aware modulation transformer and the first branch to the fourth branch, a second detection submodel is constructed;

[0086] S5-5, based on the first detection submodel and the second detection submodel, a camouflage target detection model is constructed.

[0087] As shown in Figure 7 and Figure 8As shown, the camouflage target detection model includes a second detection sub-model composed of a lightweight scale perception modulation transformer, a first branch to a fourth branch, and a first detection sub-model composed of a nonlinear dual-domain fusion module, a high-pass filter, and a low-pass filter. The first branch includes a seventh nonlinear convolutional layer, a gated edge module, a context feature enhancement module, an eighth nonlinear convolutional layer, and a second convolutional layer; the second branch to the fourth branch each include a seventh nonlinear convolutional layer, a gated edge module, a context feature enhancement module, an eighth nonlinear convolutional layer, and a second convolutional layer; the structure of each seventh nonlinear convolutional layer and each eighth nonlinear convolutional layer is the same, which is a 1x1 nonlinear convolutional layer. In Figure 8 , the first detection sub-model includes a nonlinear dual-domain fusion module, a high-pass filter, a low-pass filter, and a dual-feature fusion layer. As shown, Figure 9 , the dual-feature fusion layer includes a ninth nonlinear convolutional layer, a window division layer, a third convolutional layer, a tenth nonlinear convolutional layer, and a nonlinear multi-layer perception.

[0088] S6, processing the detection scene image by the camouflage target detection model to obtain a camouflage target detection result.

[0089] The training process of the camouflage target detection model is as follows:

[0090] S6-1, collecting original training images of a detection scene and performing data enhancement to generate detection scene training images;

[0091] Specifically, the data enhancement includes random horizontal flipping, vertical flipping, and rotation transformation, which is beneficial to enhance the generalization ability of the camouflage target detection model and effectively prevent overfitting. In this embodiment, the size of the detection scene training image is 384x384.

[0092] S6-2, inputting the detection scene training image into a lightweight scale perception modulation transformer to generate shallow multi-scale detail enhancement features , middle multi-scale semantic detail fusion features , middle scale perception global semantic association features , and deep global semantic aggregation features ; wherein, , , , The size of each of the shallow multi-scale detail enhancement features, the middle multi-scale semantic detail fusion features, the middle scale perception global semantic association features, and the deep global semantic aggregation features is 96x96x64, 48x48x128, 24x24x256, and 12x12x512, respectively.

[0093] The S6-2 includes:

[0094] S6-2-1, input the detection scene training image into the initial feature extraction layer to generate initial features; in the initial feature extraction layer, the shallow layer basic features obtained after the detection scene training image is embedded through image blocking, linear projection and preliminary convolution are taken as the initial features, which retain the original local details of the image to support subsequent multi-scale enhancement.

[0095] S6-2-2, input the initial features into the first scale perception modulation module to generate shallow multi-scale detail enhancement features;

[0096] Specifically, the initial features are processed through the multi-head mixed convolution layer, the channel dimension of the initial features is first divided into multiple independent feature heads, and then 3x3, 5x5 and other different size depth separable convolution kernels are configured for each feature head, the small size convolution kernel focuses on capturing pixel-level fine texture, and the large size convolution kernel focuses on capturing local small range structure features, and multiple groups of shallow local features of different scales are extracted through parallel calculation of multiple groups of different convolutions. Subsequently, the multiple groups of multi-scale features are transmitted into the scale perception aggregation layer, which first breaks the channel barrier between the features output by different feature heads through channel shuffling, and then integrates the dispersed multi-scale features through the grouping fusion strategy, so that the local detail information of different scales is fully interacted and complementary, and the feature redundancy is eliminated, and finally the shallow multi-scale detail enhancement features which retain the shallow rich details and fuse the multi-scale local information are output.

[0097] S6-2-3, input the shallow multi-scale detail enhancement features into the first patch embedding layer to generate the first sequence features;

[0098] Specifically, the function of the first patch embedding layer is to divide the shallow multi-scale detail enhancement features into fixed size feature patches, and convert each patch into a feature vector (Token) with uniform dimension through linear projection, while integrating position information to maintain spatial position association, to generate the first sequence features.

[0099] Through the first patch embedding layer, the two-dimensional features (shallow multi-scale detail enhancement features) can be converted into a sequence feature representation suitable for subsequent module processing, ensuring that the shallow detail information is transmitted in a structured form. By unifying the dimension and retaining the position information, it not only avoids the loss of detail features in the conversion, but also provides an adaptive input form for subsequent cross-scale feature interaction and global dependency modeling, thereby enhancing the ability to capture subtle differences of disguised targets.

[0100] S6-2-4, input the first sequence features into the second scale perception modulation module, and process them in the same way as S6-2-2 to generate middle multi-scale semantic detail fusion features;

[0101] S6-2-5, input the middle-layer multi-scale semantic detail fusion feature into the second patch embedding layer, and process it in the same way as S6-2-3 to generate a second sequence feature;

[0102] S6-2-6, input the second sequence feature into the scale-aware modulation multi-head attention module to generate a middle-layer scale-aware global semantic association feature;

[0103] Specifically, the scale-aware modulation multi-head attention module first uses a scale-aware modulation (SAM) module to perform scale-adaptive enhancement on the second sequence feature. The SAM is composed of a multi-head mixed convolution layer (MHMC) and a scale-aware aggregation layer (SAA). The former extracts multi-scale local structure information in parallel through different convolution kernels (such as 3x3, 5x5, 7x7), and the latter realizes cross-scale feature interaction through channel shuffling and grouping aggregation, thereby generating a modulation matrix M. Subsequently, the modulation matrix is multiplied element-wise with the value feature V (Z = M O V), realizing adaptive weighting of channels and space. The scale-adaptively enhanced feature output by the SAM is first projected into three groups of query, key, and value features by the MSA block, then the dot product attention weight between the query and the key is calculated and normalized by softmax, and finally the weight is used to weight and fuse the value feature element by element, so as to establish long-range dependence between features and output the integrated feature, generating the middle-layer scale-aware global semantic association feature.

[0104] S6-2-7, input the middle-layer scale-aware global semantic association feature into the third patch embedding layer, and process it in the same way as S6-2-3 to generate a third sequence feature;

[0105] S6-2-8, input the third sequence feature into the multi-head attention layer to generate a deep-layer global semantic aggregation feature. Since the multi-head attention mechanism is prior art, it will not be described in detail.

[0106] S6-3, input the shallow-layer multi-scale detail enhancement feature into the seventh nonlinear convolution layer of the first branch to generate a first convolution feature ; input the middle-layer multi-scale semantic detail fusion feature into the seventh nonlinear convolution layer of the second branch to generate a second convolution feature ; input the middle-layer scale-aware global semantic association feature into the seventh nonlinear convolution layer of the third branch to generate a third convolution feature ; input the deep-layer global semantic aggregation feature into the seventh nonlinear convolution layer of the fourth branch to generate a fourth convolution feature ;

[0107] Specifically, in order to introduce the multi-scale features output by the SMT-T into the nonlinear expression, and to strengthen the complex difference features of the camouflage target and the background, the seventh nonlinear convolution layer is arranged. The shallow multi-scale detail enhancement features are input into the seventh nonlinear convolution layer in the first branch, and the first convolution feature is generated by combining the activation function and the convolution. The seventh nonlinear convolution layer is arranged in the first branch to the fourth branch, which can refine the features and filter the redundant noise, and also adjust the feature distribution and the response range to adapt to the input requirements of the subsequent gated edge module and the context feature enhancement module. All the output features of the SMT-T need to go through this step, because these features belong to different scales, and there is a significant difference in expression ability and information density. The unified nonlinear processing can not only solve the problem of uneven expression of multi-scale features, and ensure the consistency and synergy of the subsequent parallel processing of the four branches, but also can improve the sensitivity of each scale feature to the difference information, and lay a solid foundation for the subsequent modules to fully excavate the key difference features of the camouflage target, and avoid the decline of detection accuracy due to the missing of part of the feature processing.

[0108] S6-4, input the second convolution feature to the fourth convolution feature into the first detection sub-model to generate high-frequency features and dual-domain fusion features ;

[0109] The S6-4 includes:

[0110] S6-4-1, input the second convolution feature to the fourth convolution feature into the nonlinear dual-domain fusion module to generate corresponding time-frequency global sensitive features;

[0111] Specifically, in the nonlinear dual-domain fusion module, the second convolution feature to the fourth convolution feature is processed respectively. The processing process of the second convolution feature to the fourth convolution feature is the same, so that, taking the second convolution feature as an example, the processing process of the nonlinear dual-domain fusion module is:

[0112] A1, input the second convolution feature into the spatial domain nonlinear mapping unit to generate the second spatial feature;

[0113] Specifically, in the spatial domain, the spatial domain nonlinear mapping unit is used to directly enhance the pixel-level difference. The second convolution feature is processed through the second nonlinear convolution layer, and then the second transposed convolution feature is generated by the transposed convolution layer and the first convolution layer in turn. The second convolution feature and the second transposed convolution feature are merged to increase the feature details, and the second spatial feature is generated.

[0114] A2, input the second convolution feature into the frequency domain nonlinear mapping unit to generate the second frequency domain feature;

[0115] Specifically, in the frequency domain, the amplitude spectrum and the phase spectrum are first extracted by Fourier transform through the frequency domain nonlinear mapping unit, and the amplitude (highlighting the energy difference) and the phase (strengthening the structural difference) are respectively optimized using nonlinear convolution and channel attention.

[0116] The second convolutional feature is Fourier transformed to generate a corresponding phase spectrum and amplitude spectrum. The phase spectrum is convolved through the first nonlinear convolution layer, and the key features of the data are determined through the channel attention layer to generate the phase key feature. The amplitude spectrum is processed in the same way as the phase spectrum to generate the amplitude key feature. Finally, the phase key feature, the amplitude key feature, the phase spectrum and the amplitude spectrum are merged and then inverse Fourier transformed to generate the second frequency domain feature.

[0117] A3, merge the second spatial feature and the second convolutional feature, and multiply them element by element with the second frequency domain feature to generate the second time-frequency feature;

[0118] A4, merge the second time-frequency feature and the second spatial feature to generate the corresponding time-frequency global sensitive feature.

[0119] The feature of the nonlinear dual-domain fusion module lies in its dual-domain parallel nonlinear processing, which can complement the local sensitivity of the spatial domain and the global spectral analysis of the frequency domain, thereby more efficiently mining key difference features and avoiding the limitations of single-domain processing. Finally, the enhanced features are output for subsequent high-low frequency separation.

[0120] S6-4-2, high-pass filtering and low-pass filtering are respectively performed on each time-frequency global sensitive feature to generate corresponding high-frequency features and low-frequency features ;

[0121] Specifically, a 3x3 Laplacian kernel (high-pass filtering, highlighting edge details) is used to extract the high-frequency features corresponding to the second to fourth convolutional features. A 3x3 Gaussian kernel (low-pass filtering, achieving smoothing) is used to process the time-frequency global sensitive features to obtain the low-frequency features corresponding to the second to fourth convolutional features. Then, all the high-frequency features and low-frequency features are respectively upsampled to align to the size of the second convolutional feature, and are fused through a 1x1 nonlinear convolution layer to output unified high-frequency features and low-frequency features.

[0122] High-frequency filtering can accurately capture the edge, texture and other subtle differences of the camouflage target, avoiding the loss of details. Low-frequency filtering can retain the overall structure and semantic association of the target and the background, preventing false detection by focusing only on details and ignoring the global. The unified high-frequency and low-frequency features after alignment and fusion have both the sharpness of local details and the integrity of global structure, improving the precision and robustness of camouflage target detection, especially in complex texture and multi-scale scenes.

[0123] S6-4-3, input each high-frequency feature and each low-frequency feature into a dual-feature fusion layer to generate dual-domain fusion features.

[0124] Specifically, the high-frequency feature is input into a ninth nonlinear convolutional layer for nonlinear enhancement to strengthen the expression of the detail differences of the camouflage target and generate a high-frequency convolutional feature. The low-frequency feature and the high-frequency convolutional feature are window-divided to obtain low-frequency block features and high-frequency block features. The low-frequency block features are taken as keys (K) and the high-frequency block features are taken as queries (Q). The low-frequency block features and the high-frequency block features are scaled and dot-attention weighted, and the calculation result is taken as a value (V). The weight is calculated through the scaled dot-attention mechanism to realize cross-domain information interaction of the high-frequency and low-frequency features (the low-frequency provides global structural context for the high-frequency, and the high-frequency supplements detailed positioning information for the low-frequency), and generate corresponding attention features. The attention features are input into a nonlinear multilayer perceptron to introduce nonlinear transformation to strengthen feature representation; at the same time, the original feature information is reserved through residual connection to improve robustness. Then the features are refined through a tenth nonlinear convolutional layer and a third convolutional layer in sequence to generate initial dual-domain fusion features corresponding to second to fourth convolutional features. The initial dual-domain fusion features are fused again, and finally the dual-domain fusion features are output.

[0125] Through the dual-feature fusion layer, the depth complementarity of the high-frequency and low-frequency features is realized: the global structural semantics of the low-frequency feature and the local detailed information of the high-frequency feature are fully fused through cross-attention, which anchors the high-frequency details in the global semantics and refines the low-frequency structure into precise expression with details. This fusion provides high-quality input for the subsequent global-local recognition of the decoder, effectively solving the contradiction between detail loss and global semantic ambiguity in camouflage target detection, and significantly improving the accuracy and boundary clarity of target segmentation in multi-scale and complex texture scenes.

[0126] S6-5, input the first convolutional feature and the high-frequency feature into a gated edge module to generate an edge feature map ;

[0127] Specifically, the gating weight is calculated through a first global average pooling layer and a fully connected layer, multiplied by the high-frequency feature to highlight the edge information, and finally an edge feature map is generated through a third nonlinear convolutional layer. The edge feature map is used as auxiliary injection of the context feature enhancement module. This separation fusion and edge weighting feature can retain and strengthen the detailed information, avoid loss, and at the same time closely link the edge module with the main process to improve the overall robustness.

[0128] S6-6, input the first convolutional feature and the edge feature map and the dual-domain fusion feature to a context feature enhancement module of a first branch to generate a first detail enhancement feature; input the second convolutional feature and the edge feature map and the dual-domain fusion feature to a context feature enhancement module of a second branch to generate a second detail enhancement feature; input the third convolutional feature and the edge feature map and the dual-domain fusion feature to a context feature enhancement module of a third branch to generate a third detail enhancement feature; and input the fourth convolutional feature and the edge feature map and the dual-domain fusion feature to a context feature enhancement module of a fourth branch to generate a fourth detail enhancement feature;

[0129] Specifically, since the global boundary needs frequency domain enhancement to capture long-range dependence, and the local structure needs spatial mixing to refine the pixel-level details, a spatial-frequency mixed attention is introduced into the context feature enhancement module to realize global-local interaction. The characteristic of the context feature enhancement module lies in its layer-by-layer context enhancement and cross mechanism:

[0130] The processing process of each context feature enhancement module is the same, and the input is the first / second / third / fourth convolutional feature and the edge feature map and the dual-domain fusion feature input to the corresponding context feature enhancement module. Therefore, the first context feature enhancement module is described. The processing process of the context feature enhancement module is:

[0131] S6-6-1, input the first convolutional feature to a spatial mixing unit and a second global average pooling layer respectively to generate a first spatial mixing feature and a first convolutional pooling feature;

[0132] Specifically, in the spatial mixing unit, the first convolutional feature is first processed by a first pointwise convolution layer to adjust the channel dimension. The output of the first pointwise convolution layer is processed by a fourth nonlinear convolution layer to introduce nonlinear transformation to strengthen the local detail expression. Finally, the output of the fourth nonlinear convolution layer is processed by a second pointwise convolution layer, and the output of the first pointwise convolution layer is merged to generate the first spatial mixing feature.

[0133] The application realizes fine mixing and nonlinear expression of local features in the spatial domain through channel adaptation of pointwise convolution, feature enhancement of nonlinear convolution and information reservation of residual addition, which not only maintains the basic information of the original feature, but also strengthens the detail differences (such as texture and edge) of the camouflaged target in the spatial domain, providing more discriminative spatial domain feature support for subsequent context feature enhancement.

[0134] S6-6-2, input the first convolutional feature and the dual-domain fusion feature to a mixed attention layer to generate a first mixed key feature;

[0135] Specifically, the dual-domain fusion feature is taken as a query (Q), and the first convolutional feature is taken as a key (K) and a value (V) respectively. The similarity between the query (Q) and the key (K) is calculated, and the attention weight is generated by a softmax function. The weighted feature is obtained by element-wise multiplication of the attention weight and the value (V). The weighted feature and the dual-domain fusion feature are added element by element to generate the first mixed key feature.

[0136] The application realizes the precise interaction of the local convolutional feature and the global dual-domain fusion feature by mixing the attention mechanism. The key information related to the camouflage target in the local feature is adaptively strengthened under the guidance of the global dual-domain fusion feature, while the integrity of the global semantic is preserved. This fusion not only "anchors" the local details in the global context, but also "refines" the global semantic into precise expression with details, significantly improving the discriminability of the feature to the camouflage target, and providing a more discriminative mixed feature for subsequent context enhancement and segmentation tasks.

[0137] S6-6-3, the first mixed key feature, the first spatial mixed feature and the first convolutional pooling feature are weighted and fused to be input into a fifth nonlinear convolutional layer to generate a first spatial mixed key feature;

[0138] S6-6-4, the first spatial mixed key feature and the edge feature map are merged to be input into a sixth nonlinear convolutional layer to generate a first detail enhancement feature In Figure 6 , , , respectively represent the second detail enhancement feature, the third detail enhancement feature and the fourth detail enhancement feature.

[0139] S6-7, the first detail enhancement feature is input into an eighth nonlinear convolutional layer and a second convolutional layer of the first branch to generate a shallow layer fusion feature ; the second detail enhancement feature is input into an eighth nonlinear convolutional layer and a second convolutional layer of the second branch to generate a middle layer semantic detail fusion feature ; the third detail enhancement feature is input into an eighth nonlinear convolutional layer and a second convolutional layer of the third branch to generate a middle layer global semantic fusion feature ; the fourth detail enhancement feature is input into an eighth nonlinear convolutional layer and a second convolutional layer of the fourth branch to generate a deep layer global fusion feature ;

[0140] S6-8, the shallow layer fusion feature, the middle layer semantic detail fusion feature, the middle layer global semantic fusion feature and the deep layer global fusion feature are input into a detection head to obtain a camouflage target detection training result; wherein the detection head is a nonlinear convolution sequence, including a ReLU activation function layer, a 3x3 convolution layer, a batch normalization layer and a 1x1 convolution layer.

[0141] S6-9, based on the camouflage target detection training result, adjusting the network parameters of the camouflage target detection model through deep supervision learning, completing the training of the camouflage target detection model.

[0142] To sum up, the application solves the problem of performance decline in low contrast scenes by capturing the key difference characteristics of the frequency domain amplitude spectrum and phase spectrum through the nonlinear dual-domain fusion module combined with spatial domain and frequency domain nonlinear mapping, and improves the target discrimination. Relying on the light weight scale perception modulation transformer, the dual feature fusion module extracts and aligns the multi-scale features, integrates the semantic relationship through the cross attention, and overcomes the problem of detail loss and boundary blur caused by scale diversity. The cross attention and edge auxiliary injection are integrated into the context feature enhancement module, and the multi-layer feature integration is accumulated, which takes into account the global boundary and local structure, effectively reduces the false detection and missed detection and edge roughness, and the multi-layer auxiliary supervision further strengthens the robustness.

[0143] A camouflage target detection system based on a dual-domain fusion enhanced network, comprising:

[0144] The acquisition module is used for acquiring and preprocessing the original detection scene image of the camouflage target to obtain a detection scene image.

[0145] The first construction module is used for constructing a light weight scale perception modulation transformer based on a scale perception modulation multi-head attention mechanism and a multi-head mixed convolution combined with a scale perception aggregation module.

[0146] The second construction module is used for generating a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit based on a Fourier transform, a nonlinear convolution and a channel attention mechanism, and constructing a nonlinear dual-domain fusion module.

[0147] The third construction module is used for constructing a gated edge module and a spatial mixing unit, and constructing a context feature enhancement module based on a mixed attention mechanism and a spatial mixing unit.

[0148] The model construction module is used for constructing a camouflage target detection model based on the light weight scale perception modulation transformer, the nonlinear dual-domain fusion module, the gated edge module and the context feature enhancement module.

[0149] The camouflage detection module is used for processing the detection scene image through the camouflage target detection model to obtain a camouflage target detection result.

[0150] It should be noted that, as for the system in the above embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment related to the method, and will not be described in detail here.

[0151] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0152] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0152] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting camouflaged targets based on a dual-domain fusion enhanced network, characterized in that, include: The original detection scene image of the disguised target is obtained and preprocessed to obtain the detection scene image; A lightweight scale-aware modulation transformer is constructed based on the scale-aware modulation multi-head attention mechanism and multi-head hybrid convolution, combined with a scale-aware aggregation module. Based on Fourier transform, nonlinear convolution and channel attention mechanisms, spatial domain nonlinear mapping units and frequency domain nonlinear mapping units are generated to construct a nonlinear dual-domain fusion module. Construct a gated edge module and a spatial hybrid unit; based on the hybrid attention mechanism and the spatial hybrid unit, construct a context feature enhancement module; A camouflage target detection model is constructed based on a lightweight scale-aware modulation converter, a nonlinear dual-domain fusion module, a gated edge module, and a contextual feature enhancement module. The camouflage target detection result is obtained by processing the scene image through the camouflage target detection model; The lightweight scale-aware modulation transformer includes an initial feature extraction layer, a first scale-aware modulation module, a first patch embedding layer, a second scale-aware modulation module, a second patch embedding layer, a scale-aware modulation multi-head attention module, a third patch embedding layer, and a multi-head attention layer connected in series. Both the first and second scale-aware modulation modules include multi-head hybrid convolutional layers and scale-aware aggregation layers; the scale-aware modulation multi-head attention module includes alternately stacked scale-aware modulation modules and multi-head attention blocks. The nonlinear dual-domain fusion module includes a spatial domain nonlinear mapping unit and a frequency domain nonlinear mapping unit; the frequency domain nonlinear mapping unit includes a Fourier transform layer, a first nonlinear convolutional layer, a channel attention layer, and an inverse Fourier transform layer; the spatial domain nonlinear mapping unit includes a second nonlinear convolutional layer, a transposed convolutional layer, and a first convolutional layer.

2. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 1, characterized in that, The first scale sensing modulation module and the second scale sensing modulation module have the same structure; the first patch embedding layer and the second patch embedding layer have the same structure.

3. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 1, characterized in that, The context feature enhancement module includes a hybrid attention layer, a spatial hybrid unit, a second global average pooling layer, a fifth nonlinear convolutional layer, and a sixth nonlinear convolutional layer; the spatial hybrid unit includes a first pointwise convolutional layer, a fourth nonlinear convolutional layer, and a second pointwise convolutional layer.

4. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 2, characterized in that, The construction of the camouflaged target detection model includes: The first branch is constructed based on the gated edge module, nonlinear convolutional layer, and context feature enhancement module; Based on the context feature enhancement module and the non-linear convolutional layer, the second branch, the third branch and the fourth branch are constructed respectively; A first detection sub-model is constructed based on a nonlinear dual-domain fusion module, high-pass filtering, low-pass filtering, and a nonlinear multilayer perceptron. A second detection sub-model is constructed based on a lightweight scale-aware modulation converter and the first to fourth branches. A camouflaged target detection model is constructed based on the first and second detection sub-models.

5. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 4, characterized in that, The training process of the camouflaged target detection model is as follows: Collect the original training images of the detection scene and perform data augmentation to generate the detection scene training images; The training images of the detection scene are input into a lightweight scale-aware modulation transformer to generate shallow multi-scale detail enhancement features, mid-scale multi-scale semantic detail fusion features, mid-scale scale-aware global semantic association features, and deep global semantic aggregation features. The shallow multi-scale detail enhancement features are input into the seventh non-linear convolutional layer of the first branch to generate the first convolutional feature; the mid-level multi-scale semantic detail fusion features are input into the seventh non-linear convolutional layer of the second branch to generate the second convolutional feature; the mid-level scale-aware global semantic association features are input into the seventh non-linear convolutional layer of the third branch to generate the third convolutional feature; and the deep global semantic aggregation features are input into the seventh non-linear convolutional layer of the fourth branch to generate the fourth convolutional feature. The second to fourth convolutional features are input into the first detection sub-model to generate high-frequency features and dual-domain fusion features. The first convolutional feature and high-frequency feature are input into the gated edge module to generate an edge feature map; The first convolutional feature, along with the edge feature map and the dual-domain fusion feature, is input into the context feature enhancement module of the first branch to generate the first detail enhancement feature; the second convolutional feature, along with the edge feature map and the dual-domain fusion feature, is input into the context feature enhancement module of the second branch to generate the second detail enhancement feature; the third convolutional feature, along with the edge feature map and the dual-domain fusion feature, is input into the context feature enhancement module of the third branch to generate the third detail enhancement feature; and the fourth convolutional feature, along with the edge feature map and the dual-domain fusion feature, is input into the context feature enhancement module of the fourth branch to generate the fourth detail enhancement feature. The first detail enhancement feature is input into the eighth nonlinear convolutional layer and the second convolutional layer of the first branch to generate shallow fusion features; The second detail enhancement feature is input into the eighth nonlinear convolutional layer and the second convolutional layer of the second branch to generate the mid-level semantic detail fusion feature; The third detail enhancement feature is input into the eighth nonlinear convolutional layer and the second convolutional layer of the third branch to generate the mid-level global semantic fusion feature; The fourth detail enhancement feature is input into the eighth nonlinear convolutional layer and the second convolutional layer of the fourth branch to generate deep global fusion features; The shallow fusion features, the mid-level semantic detail fusion features, the mid-level global semantic fusion features, and the deep global fusion features are input into the detection head to obtain the training results for camouflaged target detection. Based on the training results of the camouflage target detection model, the network parameters of the camouflage target detection model are adjusted through deep supervised learning.

6. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 5, characterized in that, The generation of high-frequency features and dual-domain fusion features includes: The second to fourth convolutional features are input into the nonlinear dual-domain fusion module to generate the corresponding time-frequency global sensitive features. High-pass and low-pass filtering are applied to each time-frequency global sensitive feature to generate corresponding high-frequency and low-frequency features; Each high-frequency feature and each low-frequency feature are input into the dual-feature fusion layer to generate dual-domain fusion features.

7. The method for detecting camouflaged targets based on a dual-domain fusion enhanced network according to claim 5, characterized in that, The processing procedure of the lightweight scale-sensing modulation converter is as follows: The training images of the detection scene are input into the initial feature extraction layer to generate initial features; The initial features are input into the first scale perception modulation module to generate shallow multi-scale detail enhancement features; The shallow multi-scale detail enhancement features are input into the first patch embedding layer to generate the first sequence features; The first sequence features are input into the second scale-aware modulation module to generate mid-level multi-scale semantic detail fusion features. The mid-level multi-scale semantic detail fusion features are input into the second patch embedding layer to generate the second sequence features; The second sequence features are input into the scale-aware modulation multi-head attention module to generate mid-level scale-aware global semantic association features. The mid-scale perceived global semantic association features are input into the third patch embedding layer to generate the third sequence features; The third sequence features are input into the multi-head attention layer to generate deep global semantic aggregation features.

8. A camouflaged target detection system based on a dual-domain fusion enhanced network, used to implement the camouflaged target detection method based on a dual-domain fusion enhanced network as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire the original detection scene image of the disguised target and perform preprocessing to obtain the detection scene image; The first building module is used to construct a lightweight scale-aware modulation transformer based on the scale-aware modulation multi-head attention mechanism and multi-head hybrid convolution, combined with the scale-aware aggregation module. The second building module is used to generate spatial domain nonlinear mapping units and frequency domain nonlinear mapping units based on Fourier transform, nonlinear convolution and channel attention mechanisms, and to build a nonlinear dual-domain fusion module. The third building module is used to construct the gated edge module and the spatial hybrid unit; based on the hybrid attention mechanism and the spatial hybrid unit, a context feature enhancement module is constructed. The model building module is used to construct a camouflage target detection model based on a lightweight scale-aware modulation converter, a nonlinear dual-domain fusion module, a gated edge module, and a contextual feature enhancement module. The camouflage detection module is used to process the scene image through the camouflage target detection model to obtain the camouflage target detection result.

Images