Underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative awareness

Abstract

The invention provides an underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative awareness, which comprises the following steps: inputting an underwater image to be processed into a coding and decoding network, and sampling adjacent codes by two times to obtain first to fourth scale features and bottleneck features; setting a context feature fusion module for each scale, taking the current scale and the adjacent deeper features as a pair of inputs, extracting features through global average pooling and convolution, obtaining two paths of weights through Sigmoid after splicing, and performing channel-by-channel and pixel-by-pixel weighting and aggregation on the two paths of features to form fusion features; a double-domain attention module is introduced, a context is modeled in a spatial domain, high-frequency and low-frequency components are extracted through two-dimensional discrete Fourier transform in a frequency domain and modulated through a learnable filter, and after inverse transformation, the components are fused with a spatial domain result to highlight the significance of a camouflage target and suppress a homogeneous background; and in the decoding stage, the bottleneck feature is taken as a starting point to perform step-by-step up-sampling and parallel aggregation and convolution reforming with the corresponding scale fusion feature, and finally a pixel-level segmentation result of the underwater camouflage target is output.

Description

Technical Field

[0001] This invention relates to the field of underwater camouflage target segmentation technology, and more particularly to an underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing. Background Technology

[0002] Underwater camouflage targets (such as marine organisms and biomimetic equipment) often exhibit a high degree of similarity to their background environment in color, texture, and morphology, resulting in extremely subtle differences in spectral distribution, edge gradients, and texture between the target and the background. Furthermore, the scattering and absorption of light by water further reduces imaging contrast, blurring target outlines and causing loss of detail, making identification and segmentation extremely difficult. Traditional segmentation methods based on single-scale convolutional features or single-domain enhancement struggle to effectively characterize such camouflage features with strong similarity and weak boundaries. They often face conflicts between multi-scale semantic interaction and fine-grained structural reconstruction, leading to missed detections of camouflage regions or misclassification of boundaries.

[0003] Furthermore, the salience of camouflaged targets depends on subtle differences between them and the background in the spatial and frequency domains. However, existing methods mostly model salience only in the spatial domain, lacking characterization of the high- and low-frequency energy distribution within the frequency components. The lack of coordination between high-frequency details (texture, edges) and low-frequency semantics (shape, contour) further weakens the model's sensitivity and discriminative power for weak features. In complex backgrounds, low-contrast, and non-uniform lighting scenarios, such methods struggle to achieve stable detection and accurate segmentation of camouflaged targets. Summary of the Invention

[0004] To address the problems identified in existing technologies, this invention discloses an underwater camouflage target segmentation method based on spatial and frequency domain collaborative sensing. This method introduces contextual feature fusion and a dual-domain attention mechanism into the encoding / decoding framework to jointly model the complementary relationship between spatial structure and frequency components. It captures cross-layer context and local geometric constraints in the spatial domain and analyzes high- and low-frequency energy distributions in the frequency domain, performing adaptive modulation to enhance the saliency response of the camouflage target, suppress homogeneous background interference, and achieve fine segmentation and accurate boundary recovery of weak feature regions. The specific steps of this invention are as follows: S1. Input the underwater image to be processed into a code-decode segmentation network consisting of four-level coding blocks and four-level decoding blocks. The coding features and bottleneck features of the first to fourth scales are obtained by downsampling between adjacent coding blocks by a factor of two. S2. Set up a context feature fusion module for each scale, using the encoded features of that scale as low-level features and adjacent deeper features as high-level features. The two features enter the fusion module while maintaining their original spatial resolution, without spatial alignment. S3. In the fusion module, global average pooling and 1×1 convolution are performed on the low-level and high-level features respectively to obtain the context description vector. After concatenation and Sigmoid normalization, two weights are generated. Based on the weights, the two features are weighted and aggregated one channel and one pixel at a time. S4. Further construct a dual-domain attention module to extract contextual feature responses in the spatial domain, extract high-frequency and low-frequency features in the frequency domain through two-dimensional discrete Fourier transform, and use a learnable filter to adjust the frequency band saliency. After inverse transformation, it is fused with spatial features to form a spatial-frequency co-representation. S5. In the decoding stage, the bottleneck features are used as the starting point for progressive upsampling operations. The features are then aggregated and convolved in parallel with the fusion features of the corresponding scale to gradually restore spatial details and boundaries, and finally output the pixel-level segmentation results of the underwater camouflage target.

[0005] Furthermore, the dynamic fusion of cross-scale features is achieved without spatial resampling, avoiding boundary offsets and semantic distortions caused by interpolation.

[0006] Furthermore, the context feature fusion module includes a global average pooling layer, a 1×1 convolutional layer, a concatenation layer, and a Sigmoid normalization unit. The layer and high-level features are respectively processed by global average pooling to obtain channel description vectors, and then processed by 1×1 convolution to generate compact representations. The two vectors are concatenated in the channel dimension and then input into the Sigmoid after nonlinear mapping to obtain values ​​in [0,1], which are used to generate weighting coefficients for low-level and high-level features.

[0007] Furthermore, the dual-domain attention module includes a spatial attention submodule and a frequency attention submodule; the spatial attention submodule uses max pooling and average pooling in parallel to extract the context response, and generates a spatial attention map with 3×3 convolution to highlight the structurally continuous region; the frequency attention submodule performs a two-dimensional discrete Fourier transform on the input features, decomposing them into two-way features: high-frequency components (edges, textures) and low-frequency components (shapes, contours).

[0008] Furthermore, the high and low frequency components of the frequency attention submodule are adjusted by a learnable filter to generate a corresponding frequency band mask to enhance high frequencies and suppress redundant low frequencies; then, they are reconstructed into frequency domain enhancement features through inverse two-dimensional discrete Fourier transform, and short-circuit residuals are retained to avoid detail distortion caused by frequency domain reconstruction.

[0009] Furthermore, the spatial and frequency domain outputs are concatenated in the channel dimension and then fused by a 3×3 convolution. A dual-domain weight mask is generated by using Sigmoid. This mask modulates the fused features element-wise in both the channel and spatial dimensions, thereby adaptively improving the significant response and suppressing noise interference in weak contrast and homogeneous background regions.

[0010] Furthermore, the decoding stage adopts a step-by-step upsampling and cross-layer aggregation approach. The upsampling results at each level and the fusion features at the corresponding scale are convolved and reconstituted to achieve detail restoration and boundary enhancement.

[0011] Furthermore, the bottleneck layer performs channel compression through 1×1 convolution and then feature reshaping through 3×3 convolution to improve the feature transfer efficiency and robustness in the decoding stage.

[0012] Furthermore, the method achieves significant enhancement and accurate boundary recovery of camouflaged targets in complex backgrounds through a spatial domain-frequency domain collaborative mechanism. The spatial domain is responsible for global semantic and local geometric consistency modeling, while the frequency domain performs adaptive modulation for high-frequency details and low-frequency contours. The two work together under the action of dual-domain weighted masking, effectively suppressing homogeneous interference and noise false detections while maintaining the integrity of the target structure, thereby obtaining stable and reliable pixel-level segmentation results.

[0013] Compared with the prior art, the present invention has the following advantages: 1. This invention adopts a spatial domain-frequency domain collaborative perception mechanism, introducing dual-domain attention on the basis of context fusion. High-level features are decomposed into high-frequency and low-frequency components by two-dimensional discrete Fourier transform, and frequency band selective enhancement is performed through a learnable filter. After inverse transformation, they are fused with spatial domain features according to weights, which can simultaneously enhance high-frequency edges and textures as well as low-frequency shapes and contours. Under the conditions of homogeneous background and weak contrast, it can effectively highlight camouflaged targets, reduce adhesion misclassification and boundary loss, and ensure contour continuity and structural integrity.

[0014] 2. The context feature fusion module adopts the native scale weight generation method, with the current scale as the low layer and the adjacent deeper layers as the high layer. Global average pooling and 1×1 convolution are performed within their respective scales, and two weights are obtained by passing through Sigmoid. The two features are weighted channel-wise and pixel-wise respectively. The aggregation end only performs online sampling on the low-level branch to match the output scale, realizing the generation of weights at the same scale and minimum alignment fusion, avoiding interpolation distortion and boundary offset in the weight calculation stage, and taking into account both fine-grained texture fidelity and global semantic consistency.

[0015] 3. The decoding end adopts a depth-order reconstruction strategy, which starts with the bottleneck feature and upsamples it step by step. It also aggregates and reconstructs the dual-domain fusion features at each scale in parallel. The path is mainly composed of lightweight units, which can enhance detail filling and boundary restoration while suppressing noise backflow and homogeneous background interference, thus achieving a balance between high-precision segmentation and low inference overhead.

[0016] In summary, the technical solution of this invention can significantly improve the segmentation accuracy and robustness in underwater camouflage scenarios, addressing the issues of strong similarity and weak boundaries. It can simultaneously obtain segmentation results with structural integrity and accurate boundaries in complex backgrounds and low-contrast environments, demonstrating good promotion and application value.

[0017] Based on the above reasons, this invention can be widely applied in related technical fields such as underwater visual perception and intelligent operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing of the present invention.

[0020] Figure 2 This is a schematic diagram of the network structure of the underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing of the present invention.

[0021] Figure 3 This is a schematic diagram of the network structure of the context feature fusion module in the underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing of the present invention.

[0022] Figure 4 This is a schematic diagram of the dual-domain attention module network structure in the underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative perception of the present invention.

[0023] Figure 5 This is a schematic diagram comparing the underwater camouflage target segmentation result with the actual underwater camouflage target label using spatial domain and frequency domain collaborative sensing in an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] like Figure 1 As shown, this invention provides an underwater camouflage target segmentation method based on spatial domain-frequency domain collaborative sensing, characterized by comprising: S1. Input the underwater image to be processed into a code-decode segmentation network consisting of four-level coding blocks and four-level decoding blocks. The coding features and bottleneck features of the first to fourth scales are obtained by downsampling between adjacent coding blocks by a factor of two. In a specific implementation, as a preferred embodiment of the present invention, the bottleneck layer is compressed through 1×1 convolution and then reorganized through 3×3 convolution to improve the feature transfer efficiency and robustness in the decoding stage.

[0027] S2. Set up a context feature fusion module for each scale, using the encoded features of that scale as low-level features and adjacent deeper features as high-level features. The two features enter the fusion module while maintaining their original spatial resolution, without spatial alignment. In a specific implementation, as a preferred embodiment of the present invention, the dynamic fusion of cross-scale features is achieved without spatial resampling, thereby avoiding boundary offsets and semantic distortions caused by interpolation.

[0028] S3. In the fusion module, global average pooling and 1×1 convolution are performed on the low-level and high-level features respectively to obtain the context description vector. After concatenation and Sigmoid normalization, two weights are generated. Based on the weights, the two features are weighted and aggregated one channel and one pixel at a time. In a specific implementation, as a preferred embodiment of the present invention, the context feature fusion module includes a global average pooling layer, a 1×1 convolutional layer, a concatenation layer, and a Sigmoid normalization unit. The layer and high-level features are respectively processed by global average pooling to obtain channel description vectors, and then processed by 1×1 convolution to generate compact representations. The two vectors are concatenated in the channel dimension and input into the Sigmoid after nonlinear mapping to obtain values ​​in [0,1], which are used to generate weighting coefficients for low-level and high-level features.

[0029] S4. Further construct a dual-domain attention module to extract contextual feature responses in the spatial domain, extract high-frequency and low-frequency features in the frequency domain through two-dimensional discrete Fourier transform, and use a learnable filter to adjust the frequency band saliency. After inverse transformation, it is fused with spatial features to form a spatial-frequency co-representation. In a specific implementation, as a preferred embodiment of the present invention, the dual-domain attention module includes a spatial attention submodule and a frequency attention submodule; the spatial attention submodule uses max pooling and average pooling in parallel to extract the context response, and generates a spatial attention map with 3×3 convolution to highlight the structurally continuous region; the frequency attention submodule performs a two-dimensional discrete Fourier transform on the input features, decomposing them into two paths of features: high-frequency components (edges, textures) and low-frequency components (shapes, contours).

[0030] The high and low frequency components of the frequency attention submodule are adjusted by a learnable filter to generate a corresponding frequency band mask to enhance high frequencies and suppress redundant low frequencies. Subsequently, the frequency domain enhancement features are reconstructed through inverse two-dimensional discrete Fourier transform, and the short-circuit residual is retained to avoid detail distortion caused by frequency domain reconstruction.

[0031] The spatial and frequency domain outputs are concatenated in the channel dimension and then fused by a 3×3 convolution. A dual-domain weight mask is generated by Sigmoid. This mask modulates the fused features element-wise in both the channel and spatial dimensions, thereby adaptively improving the significant response and suppressing noise interference in weak contrast and homogeneous background regions.

[0032] S5. In the decoding stage, the bottleneck features are upsampled step by step by 2, and then aggregated and convolved in parallel with the fusion features of the corresponding scale to gradually restore spatial details and boundaries, and finally output the pixel-level segmentation results of the underwater camouflage target.

[0033] In a specific implementation, as a preferred embodiment of the present invention, the decoding stage adopts a step-by-step upsampling and cross-layer aggregation method. The upsampling results of each level and the fusion features of the corresponding scale are convolved and reorganized before being spliced ​​together to achieve detail restoration and boundary enhancement.

[0034] The proposed method achieves significant enhancement and accurate boundary recovery of camouflaged targets in complex backgrounds through a spatial-frequency domain collaborative mechanism. The spatial domain is responsible for global semantic and local geometric consistency modeling, while the frequency domain performs adaptive modulation for high-frequency details and low-frequency contours. The two work together under the action of dual-domain weighted masking, effectively suppressing homogeneous interference and noise false detections while maintaining the integrity of the target structure, thereby obtaining stable and reliable pixel-level segmentation results.

[0035] Example like Figure 2As shown, the underwater image to be processed is input into a encoder-decoder segmentation network consisting of four levels of encoder blocks and four levels of decoder blocks. Adjacent encoder blocks are downsampled by a factor of two to obtain the encoded features at the first to fourth scales sequentially. ~ And bottleneck characteristic B.

[0036] like Figure 3 As shown, a context feature fusion module is set up for each scale. The encoded features of that scale are used as low-level features and the adjacent deeper features are used as high-level features. The two features enter the fusion module while maintaining their original spatial resolution, without spatial alignment. Secondly, in the fusion module, global average pooling and 1×1 convolution are performed on the low-level and high-level features respectively to obtain the context description vector. After concatenation and Sigmoid normalization, two weights are generated, and the two features are weighted and aggregated on a channel-by-channel and pixel-by-pixel basis according to the weights. like Figure 4 As shown, a dual-domain attention module is further constructed to extract contextual feature responses in the spatial domain and extract high-frequency and low-frequency features in the frequency domain through two-dimensional discrete Fourier transform. The frequency band saliency is adjusted by using a learnable filter, and after inverse transformation, it is fused with spatial features to form a spatial-frequency co-representation. Finally, the decoding stage starts with the bottleneck features and upsamples them by 2 at each level. These features are then aggregated and reconvolved in parallel with the fusion features at the corresponding scales to gradually restore spatial details and boundaries, ultimately outputting pixel-level segmentation results for underwater camouflage targets.

[0037] like Figure 5 As shown in the figure, this embodiment presents a comparison diagram between the underwater camouflage target segmentation results based on spatial domain-frequency domain collaborative sensing and the real labels. The first column is the original underwater image, the second column is the corresponding manually annotated real labels, and the third column is the segmentation prediction mask of the method of this invention. As can be seen from the figure, in complex camouflage environments such as algae blooms, rocky reef backgrounds, sponge attachment, and low illumination, the method can still maintain the integrity and continuity of the target boundary, significantly suppress the adhesion between the target and the background, and effectively restore the fine-grained structure of occluded or weakly contrasted areas. The prediction results are highly consistent with the real labels in terms of shape, boundary, and region consistency, fully demonstrating the robustness and high-precision segmentation capability of this invention in diverse underwater camouflage scenarios.

[0038] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0039] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0040] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0041] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0042] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for underwater camouflage target segmentation based on spatial domain and frequency domain collaborative sensing, characterized in that: The underwater image to be processed is acquired, and a coding-decoding segmentation network consisting of four-level coding blocks and four-level decoding blocks is constructed. The coding features and bottleneck features of the underwater image at the first to fourth scales are obtained by downsampling between adjacent coding blocks by a factor of 2. For each scale of encoded features, a corresponding context feature fusion module is set up. The encoded features at this scale are used as low-level features and adjacent deeper features are used as high-level features. The two features are entered into the context feature fusion module with their original spatial resolution intact, without spatial alignment. In the context feature fusion module, global average pooling and 1×1 convolution are performed on low-level features and high-level features respectively to obtain context description vectors. After concatenation and Sigmoid normalization, two weights are generated. Based on these weights, the two features are weighted and aggregated channel-wise and pixel-wise to form fused features. A dual-domain attention module is constructed to extract the contextual feature response of underwater images in the spatial domain and extract high-frequency and low-frequency features in the frequency domain through two-dimensional discrete Fourier transform. The frequency band saliency is adjusted by a learnable filter and then fused with spatial features after inverse transform to form a spatial-frequency co-representation. The decoding stage starts with bottleneck features and performs upsampling step by step, and then aggregates and reconstructs them in parallel with the fusion features of the corresponding scale to gradually restore spatial details and boundaries. The final output is pixel-level segmentation results of the underwater camouflaged target.

2. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The context feature fusion module includes a global average pooling layer, a 1×1 convolutional layer, a concatenation layer, and a Sigmoid normalization unit, which are used to generate weighted coefficients for low-level features and high-level features.

3. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The dual-domain attention module includes a spatial attention submodule and a frequency attention submodule. The spatial attention submodule uses max pooling and average pooling in parallel to extract contextual features of the underwater image, while the frequency attention submodule decomposes the input features into high-frequency and low-frequency components through two-dimensional discrete Fourier transform.

4. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The high and low frequency components of the frequency attention submodule are adjusted by a learnable filter and then reconstructed into frequency domain enhancement features through inverse two-dimensional discrete Fourier transform.

5. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The spatial and frequency domain outputs are convolved and fused, and a dual-domain weight mask is generated using Sigmoid to dynamically adjust the response of significant regions in both the channel and spatial dimensions.

6. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The decoding stage employs a step-by-step upsampling method and a cross-layer aggregation method. The upsampling results at each level are convolved with the fusion features at the corresponding scale and then stitched together to achieve detail restoration and boundary enhancement.

7. The underwater camouflage target segmentation method based on spatial domain and frequency domain collaborative sensing according to claim 1, characterized in that, The bottleneck layer performs channel compression through convolution, followed by convolutional feature reorganization, to improve the efficiency and robustness of feature transfer during the decoding stage.

Images