A frequency domain-space domain collaborative perception method for underwater salient target detection

Abstract

This application belongs to the field of computer vision technology, specifically relating to an underwater salient target detection method based on frequency-spatial domain collaborative perception. The method includes a frequency-spatial domain collaborative perception module, a cross-scale bidirectional interaction module, and a deformable attention module guided by a reverse region-boundary approach. In the encoding stage, given an input image, a four-layer feature map is first extracted through a backbone network. The extracted feature maps are then compressed through convolutional layers. The convolutional feature maps are then processed by the frequency-spatial domain collaborative perception module, effectively suppressing color distortion in the underwater environment. In the decoding stage, to further improve the model's target localization accuracy, these features are integrated with multi-scale contextual information through the cross-scale bidirectional interaction module via cross-scale convolutional paths, enhancing the target's spatial structure perception capability. Finally, the enhanced target features are fed into the deformable attention module guided by a reverse region-boundary approach for fine-tuning and boundary optimization of the underwater target region.

Description

Technical Field

[0001] This application belongs to the field of computer vision technology, specifically relating to a method for detecting salient underwater targets using frequency-spatial domain collaborative sensing. Background Technology

[0002] Early underwater salient target detection methods primarily relied on manually designed features, such as color contrast, texture differences, or prior saliency assumptions. These methods typically relied on empirical modeling of local statistical properties of underwater images and achieved some success under relatively simple imaging conditions. However, due to the drastic changes in lighting conditions and diverse degradation patterns in underwater scenes, manually designed features struggle to fully describe the complex degradation characteristics, leading to inaccurate target-background separation. Furthermore, the prediction results were highly sensitive to noise, and boundaries were prone to blurring or breakage. With the development of deep learning technology, underwater salient target detection methods based on deep neural networks have gradually become mainstream. Existing deep learning methods are mainly improved in the following directions: First, by designing lightweight network structures and efficient operators, the number of model parameters and computational complexity are reduced to meet the real-time requirements of practical deployments; second, by introducing boundary or structure enhancement mechanisms, target contour information is explicitly modeled to improve the boundary accuracy of detection results; third, by fusing auxiliary modal information such as depth and jointly modeling with RGB images, robustness under complex backgrounds and severe color shift conditions is enhanced; and fourth, physical priors of underwater imaging are introduced into the network learning process to improve the interpretability and cross-scene generalization ability of the model.

[0003] While the aforementioned methods have improved underwater saliency detection performance to some extent, they still have significant shortcomings. On the one hand, lightweight methods, while reducing computational overhead, often sacrifice feature representation capabilities, leading to missed detections and incomplete target information in severely degraded scenarios. On the other hand, boundary enhancement methods are highly dependent on the quality of boundary cues; however, under strong underwater scattering and low contrast conditions, boundary information is often blurred or discontinuous, easily introducing noise and affecting detection stability. Multimodal fusion methods are limited by the availability and quality of auxiliary modal data; their performance significantly degrades when depth information is inaccurate or missing, limiting their practical application. Although physically guided methods enhance the model's generalization ability to some extent, their modeling process is complex, and in the absence of clear real references, the generated results may deviate from the true appearance of the target. Further analysis reveals that the core modeling paradigm of existing underwater saliency target detection methods still mainly relies on spatial domain feature representation, limiting their perception range to local areas and making it difficult to effectively address the global degradation problem prevalent in underwater imaging. Summary of the Invention

[0004] To address the aforementioned problems, this application provides a frequency-spatial domain collaborative sensing method for underwater salient target detection, aiming to simultaneously improve the accuracy of target localization, the integrity of salient regions, and the fineness of target boundaries in complex underwater environments. The technical solution is as follows: A frequency-spatial domain collaborative sensing method for underwater salient target detection includes a frequency-spatial domain collaborative sensing module, a cross-scale bidirectional interaction module, and a reverse region-boundary guided deformable attention module; The frequency-space domain collaborative sensing module includes an adaptive frequency sensing module and a frequency-space domain interactive fusion module; Adaptive frequency sensing module: extracts low-frequency and high-frequency components; low-frequency components are used to enhance global structural information and model the response of underwater target, while high-frequency components focus on enhancing edge details that are blurred due to turbidity and noise. Frequency-spatial domain interactive fusion module: It uses correlation guidance and bidirectional gating mechanism to allow frequency-spatial features to mutually correct and enhance each other during interaction. Spatial details are used to suppress noise uncertainty areas in the frequency domain response. The global structural prior provided by the frequency domain inversely constrains the perception range of spatial features, guiding them to focus on the underwater target area. Cross-scale bidirectional interaction module: Fine-grained edge information of small-scale features is effectively combined with semantic expression of large-scale features to enhance the saliency response of targets at different scales; Reverse Region-Boundary Guided Deformable Attention Module: Introducing "reverse region-reverse boundary" as a structural constraint guides deformable convolution for dynamic sampling, thereby achieving collaborative modeling of region integrity and boundary refinement.

[0005] Preferably, the adaptive frequency sensing module achieves adaptive decoupling between global structure preservation and local detail enhancement according to different underwater degradation modes, specifically as follows: The input image is processed by a convolutional layer to extract low-dimensional feature representations, and a high-level semantic guidance mechanism is introduced to multiply the high-level features with position awareness with the low-level features element by element. (1); in The element-wise multiplication operation is represented by DP(a,b) / UP(a,b), which means adjusting the spatial dimension of feature map a to the spatial dimension of feature map b. The convolutional feature map is denoted as... .

[0006] Preferably, the adaptive frequency sensing module is equipped with a dynamic threshold and frequency masking mechanism, specifically as follows: 1) Compressed contextual representations are obtained through global average pooling to retain key semantic information and perceive overall scene characteristics; 2) Construct a lightweight prediction module consisting of two convolutional layers and nonlinear activation to perform nonlinear mapping on features; 3) Introduce Sigmoid activation to generate two normalized scaling coefficients. and It is used to adaptively adjust the polar coordinate radius threshold, thereby dynamically dividing low-frequency and high-frequency components; (2); (3); Among them, Cov, GAP and These represent convolutional layers, global average pooling, and the sigmoid activation function, respectively. Rlow represents the low-frequency radius, and H is the image height. 4) After obtaining the low-frequency radius, two indicator function masks were constructed based on the polar coordinate grid to distinguish between low-frequency and high-frequency components. The indicator functions for the low-frequency and high-frequency regions are defined as follows: (4); (5); (6); Where, x grid and y grid Let I represent the normalized grid coordinates in the horizontal and vertical directions, respectively, with a value range of [-1, 1]. A (x) represents an indicator function on set A, which takes the value 1 if and only if x∈A; 5) After obtaining the high-frequency and low-frequency masks, process the feature map. Frequency domain features are obtained by performing a fast Fourier transform. Then, based on the low-frequency mask and the high-frequency mask, Enhancement is performed, and finally, the low-frequency and high-frequency components are transformed back from the frequency domain to the spatial domain using inverse Fourier transform to obtain separated low-frequency feature maps. and high-frequency feature maps ; 6) Introduce hyperparameters This is used to control the relative proportions of low-frequency and high-frequency components, thereby obtaining the final output. : (9).

[0007] Preferably, the frequency domain-spatial domain interactive fusion module employs an attention mechanism, including channel attention and spatial attention. Channel attention adaptively adjusts the weights of different feature channels to highlight channel features that contribute to saliency discrimination and suppress invalid channels affected by color distortion or noise. Spatial attention enhances the response of salient regions while weakening complex background regions, helping the network to more accurately locate underwater targets. The expression is as follows: (10); (11); (12); (13); Where CA represents channel attention and SA represents spatial attention.

[0008] Preferably, cross-modal correlation guidance is introduced during the fusion process. By calculating the similarity factor between spatial and frequency domain features, significant regions consistent across modalities are highlighted, and spurious regions in single-modal features are effectively suppressed. The expression is as follows: (14); (15); (16); (17); C represents the number of channels.

[0009] Preferably, a bidirectional gating mechanism is designed, which modulates the frequency domain features with spatial features to enhance their boundary and texture details; at the same time, the frequency domain features are used to guide the spatial features to improve global consistency, and the original modal features are introduced into the fusion result in the form of residuals, as shown in the following expression: (18); (19); (20); (twenty one); (twenty two); (twenty three); (twenty four); Finally, a channel weighting factor w is introduced to obtain the adaptive fusion result: (25); Where w represents the following: (26); Where CRCS represents convolution + ReLU activation function + convolution + Sigmoid activation function.

[0010] Preferably, the cross-scale bidirectional interaction module includes multi-scale convolutional feature extraction, cross-scale bidirectional interaction, and weighted fusion; Multi-scale convolutional feature extraction includes three branches of different scales. Each branch extracts features through convolution kernels of different sizes. The feature maps of each branch are stitched together to form a fused feature map. Cross-scale bidirectional interaction comprises two parts: channel interaction and spatial interaction. Large-scale features are used to generate channel attention, guiding the channel selection of small-scale features; small-scale features are used to generate spatial attention to correct the spatial response of large-scale features, as shown in the following expression: (30); (31); (32); For small-medium ( , ), medium-large ( , ), small-large ( , The three scales were interacted with to obtain three pairs of enhanced features. , and ; The final fused features are then convolved to obtain an enhanced output: (33)

[0011] Preferably, the reverse region-boundary guided deformable attention module includes a feature alignment module, wherein the specific steps of feature alignment are: upsampling the decoded features... Adjust to integrate features with the coding stage Using the same spatial resolution, and after refinement processes such as stitching and convolution, batch normalization, and ReLU activation, uniform fused features are obtained. .

[0012] Preferably, the reverse region-boundary guided deformable attention module includes a reverse region deformable attention module. This module no longer relies on positive saliency regions as the primary constraint, but instead explicitly models the background region through a reverse region mask. The specific steps are as follows: First, a background mask image is generated through thresholding. Secondly, the fused features are multiplied element-wise with the background mask image; Finally, guided by the background mask, a mask-aware deformable convolution is introduced. This convolution can adaptively adjust the sampling position, ensuring the continuity of the overall target structure and the distinction from the background. (36); (37); ; (38); in As the background mask image, This represents the weight of the convolution kernel at position p. denoted by , where p represents the center position of the convolution kernel, and p represents the position of a fixed sampling point within the convolution kernel relative to the center. This represents the offset generation function. It is the offset vector of each convolution sampling point.

[0013] Preferably, the reverse region-boundary guided deformable attention module includes reverse boundary deformable attention, which treats boundary regions as unstable factors for explicit suppression, while utilizing non-boundary regions to provide a more stable structural reference, as follows: 1) Underwater salient map Gradient calculations are performed to obtain edge responses in both the horizontal and vertical directions, and non-maximum suppression is combined to generate an initial boundary map. : 2) To achieve reverse boundary guidance, the fused features will be... With reverse boundary map Element-by-element multiplication yields reliable structural features. ; 3) Introducing boundary-aware deformable convolution, which modulates the sampling offset through boundary response, enabling the convolution kernel to generate adaptive displacement near the edge, thus realizing a dynamic sampling process for contour awareness: (42); ; (43)

[0014] Preferably, the reverse region-boundary guided deformable attention module includes a region-boundary collaborative fusion module, which fuses the reverse region deformable attention output with the reverse boundary deformable attention output to generate a refined saliency prediction result, as shown in the following expression: (44)

[0015] Compared with the prior art, the beneficial effects of this application are as follows: The frequency-space domain collaborative sensing module explicitly models the compensation and constraint relationship between global consistency in the frequency domain and local discrimination in the spatial domain through a learnable frequency adaptive mechanism and cross-domain bidirectional gating interaction, thereby specifically alleviating the problems of inaccurate global target structure and loss of boundary details caused by underwater degradation.

[0016] The cross-scale bidirectional interaction module achieves adaptive perception of targets at different scales through a bidirectional information flow mechanism.

[0017] The reverse region-boundary guided deformable attention module aims to improve the model's structural perception capability in underwater degradation scenarios from two complementary levels: region integrity and boundary refinement. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the present invention; Figure 2 This is a schematic diagram of the adaptive frequency sensing module; Figure 3 This is a schematic diagram of the frequency domain-spatial domain interactive fusion module; Figure 4 A visual comparison and analysis chart of feature maps at different stages; Figure 5 A diagram of a cross-scale, two-way interactive module; Figure 6 A schematic diagram of a deformable attention module guided by a reverse region-boundary; Figure 7 This is a saliency map visualization comparison between the present invention and 10 more advanced saliency target detection methods. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1This paper presents a frequency-spatial domain collaborative sensing method for underwater salient target detection. It follows an encoder-decoder design structure specifically designed for underwater salient target detection scenarios, aiming to alleviate the problem of inaccurate saliency map prediction caused by complex underwater imaging environments. Specifically, in the encoding stage, given an input image... First, four feature maps are extracted using the backbone network PVTv2. Fp i ,in i ∈{1,2,3,4}, the spatial resolution of each feature map layer is as follows: W / 2 i+1 ×H / 2 i+1 Where W is the image width and H is the image height. To reduce the number of parameters and improve computational efficiency, the extracted feature maps are first compressed through a convolutional layer. The convolutional feature map is denoted as... Fspa i The number of channels is C i ∈{64}. Next, these features pass through a frequency-spatial domain collaborative perception module, which effectively suppresses color distortion, turbidity, and background noise interference in the underwater environment, thereby enhancing the discriminative ability of target features. In the decoding stage, to further improve the model's target localization accuracy, these features pass through a cross-scale bidirectional interaction module, which integrates multi-scale contextual information through cross-scale convolutional paths, strengthening the spatial structure perception ability of the target. Finally, the enhanced target features enter a reverse region-boundary guided deformable attention module for fine-tuning and boundary optimization of the underwater target region.

[0021] The specific steps are as follows: B. Frequency-spatial domain collaborative sensing module: In complex underwater environments, color distortion, water turbidity, and background clutter often weaken the global structural response and blur the local boundary representation of prominent targets, severely limiting the stability and accuracy of detection models. Existing methods typically face a dilemma: while spatial domain methods can characterize details, their global representation is prone to instability under strong degradation; frequency domain methods can provide robust global structural priors but lack the ability to accurately model local details. To address this, this invention proposes an adaptive frequency-spatial domain collaborative sensing module for underwater degradation. Its core lies in explicitly modeling the compensation and constraint relationship between global consistency in the frequency domain and local discriminative power in the spatial domain through a learnable frequency adaptation mechanism and cross-domain bidirectional gating interaction, thereby specifically mitigating the problems of inaccurate global target structure and loss of boundary details caused by underwater degradation. This module mainly consists of two parts: an adaptive frequency sensing module and a frequency-spatial domain interactive fusion module. The adaptive frequency sensing module employs a learnable dynamic threshold and frequency masking mechanism. Low-frequency components enhance global structural information and model the underwater target's response, while high-frequency components focus on reinforcing edge details blurred by turbidity and noise. This combination of high and low-frequency features constructs a more discriminative frequency domain feature set for underwater scenes. The frequency-spatial domain interaction fusion module addresses the advantage cancellation problem caused by simple cross-domain feature fusion by utilizing correlation guidance and a bidirectional gating interaction mechanism to allow the two features to mutually correct and enhance each other during interaction. Spatial details are used to suppress noise uncertainty areas in the frequency domain response. Simultaneously, the global structural prior provided by the frequency domain inversely constrains the perception range of the spatial features, guiding them to focus on the underwater target area. This bidirectional constraint effectively achieves synergistic optimization of noise suppression and target enhancement, ultimately significantly improving the consistency of target localization even under strong degradation conditions. The structure and working mechanism of these two modules will be explained in detail below.

[0022] Adaptive frequency sensing module: The designed adaptive frequency sensing module, such as Figure 2 As shown, unlike existing methods that only statically decompose or uniformly weight frequency domain features, this module explicitly models frequency components as discriminative units. Through a learnable dynamic threshold and frequency masking mechanism, it adaptively selects and suppresses the effectiveness of each frequency component under the current degradation conditions, thereby enhancing underwater saliency-related features at different frequency domain levels. Specifically, this module adaptively separates low-frequency and high-frequency components based on input features: low-frequency components compensate for the weakening of the global structural response caused by underwater degradation, maintaining the overall consistency of the target subject region; high-frequency components address edge blurring caused by turbidity and noise, enhancing edge details that contribute to saliency discrimination. This frequency-discrimination-driven dynamic separation strategy achieves adaptive decoupling between global structure preservation and local detail enhancement according to different underwater degradation modes, thus constructing a more robust and discriminative frequency domain representation foundation for complex underwater scenes.

[0023] First, the input image is processed by convolutional layers to extract low-dimensional feature representations. However, due to underwater image degradation, these low-dimensional features are often accompanied by significant background noise and blur artifacts, making it difficult to accurately represent salient targets. To address this, this invention introduces a high-level semantic guidance mechanism, which performs element-wise multiplication of high-level features with location awareness with low-level features. This operation fully utilizes the prior location information provided by the high-level features, effectively highlighting the target region in the low-dimensional features, thereby achieving structural enhancement and region focusing for salient underwater targets.

[0024] (1); in This indicates the element-wise multiplication operation. DP(a,b) / UP(a,b) Indicates the feature map a Spatial dimension adjusted to feature map b Spatial dimension.

[0025] To more effectively distinguish between low-frequency and high-frequency features in complex underwater degradation environments, this invention proposes a dynamic threshold generation mechanism based on polar coordinates. This mechanism adaptively determines the boundary between low-frequency and high-frequency components, enabling the network to adaptively balance global structure preservation and local detail enhancement, thereby alleviating the insufficient adaptability of traditional frequency domain decomposition based on fixed frequency thresholds in underwater degradation scenarios. Specifically, firstly, the module obtains compressed contextual representations through global average pooling to retain key semantic information and perceive overall scene characteristics; subsequently, a lightweight prediction module consisting of two convolutional layers and nonlinear activation is constructed to perform nonlinear mapping on the features; finally, a dynamic threshold generation mechanism is introduced... Sigmoid Activate the generation of two normalized scaling coefficients β 1 and β 2. Used to adaptively adjust the polar coordinate radius threshold, thereby dynamically dividing low-frequency and high-frequency components.

[0026] (2); (3); in Cov , GAP and These represent convolutional layers, global average pooling, and... Sigmoid Activation function R low Indicates the low-frequency radius. After obtaining the low-frequency radius, two indicator function masks were constructed based on a polar coordinate grid to distinguish between low-frequency and high-frequency components. The indicator functions for the low-frequency and high-frequency regions are defined as follows: (4); (5); (6); in, x grid and y grid These represent the normalized grid coordinates in the horizontal and vertical directions, respectively, with a value range of [-1, 1]. A ( x ) represents a set A The indicator function on, if and only if x ∈ A The value is 1.

[0027] After obtaining the high-frequency and low-frequency masks, the feature map is processed. Fen i Frequency domain features are obtained by performing a fast Fourier transform. Then, based on the low-frequency mask and the high-frequency mask, Enhancement is performed, and finally, the low-frequency and high-frequency components are transformed back from the frequency domain to the spatial domain using inverse Fourier transform to obtain separated low-frequency and high-frequency feature maps: (7); (8); in and These represent the Fourier transform and the inverse Fourier transform, respectively.

[0028] Finally, hyperparameters are introduced. This is used to control the relative proportions of low-frequency and high-frequency components, thereby obtaining the final output: (9).

[0029] Frequency domain – spatial domain interactive fusion module: In the adaptive frequency perception module, high-frequency components are selectively enhanced to highlight edges and details that contribute to saliency discrimination while suppressing noise interference. However, frequency domain features still primarily focus on global semantic modeling, with relatively weak boundary localization and local texture characterization. In contrast, spatial domain features can provide more accurate boundary and local texture information. To fully utilize the global consistency of the frequency domain and the local discriminative power of the spatial domain, this invention designs a frequency-spatial domain interactive fusion module. Through attention mechanisms, relevance guidance, and gating interaction strategies, it achieves deep interaction and adaptive fusion of the two types of features. During the fusion process, spatial domain features are used to repair local structures affected by underwater degradation, while frequency domain global features provide overall layout constraints, guiding the network to focus on salient target areas, thereby improving the accuracy of boundary and detail characterization. The overall structure of this module is as follows: Figure 3 As shown, the specific details are as follows.

[0030] (1) Single-modal attention enhancement: Before cross-modal fusion, it is necessary to ensure that the feature representation of each modality is sufficient and robust. To this end, this invention introduces channel attention and spatial attention within each modality. Channel attention adaptively adjusts the weights of different feature channels to highlight channel features that contribute to saliency discrimination and suppress invalid channels affected by color distortion or noise. Spatial attention enhances the response of salient regions while weakening complex background regions, helping the network to more accurately locate underwater targets. Through this single-modal attention enhancement, the network can obtain more robust and discriminative features in complex underwater scenes.

[0031] (10); (11); (12); (13); in CA Indicates channel attention. SA This represents spatial attention.

[0032] (2) Correlation Guidance: Underwater targets are easily affected by background interference, which can easily lead to false detections. To alleviate this problem, this invention introduces cross-modal correlation guidance during the fusion process. By calculating the similarity factor between spatial and frequency domain features, it highlights salient regions that are consistent across modes and effectively suppresses spurious regions in a single mode. This strategy enhances the model's ability to distinguish true salient regions, thereby effectively reducing background interference.

[0033] (14); (15); (16); (17).

[0034] (3) Gated Interactive Fusion: A major challenge in underwater saliency detection lies in the inconsistency between modalities: spatial domain features can provide clear boundary and local texture information, but under the interference of water turbidity and color distortion, the global structure and overall target localization ability are insufficient, which can easily lead to incomplete target segmentation; while frequency domain features can maintain robustness to underwater color distortion and water turbidity through global frequency modeling, but its global modeling tends to result in insufficient depiction of boundaries and local details, making it impossible to accurately recover target edges and local textures. Direct splicing or weighted fusion can easily lead to conflicts. To this end, this invention designs a bidirectional gating mechanism, which modulates frequency domain features with spatial features to enhance their boundary and texture details; at the same time, it uses frequency domain features to guide spatial features to improve global consistency. This complementary interaction effectively alleviates modal differences. In the fusion process, in order to avoid the loss of original information, this invention introduces the original modal features into the fusion result in the form of residuals, making the fused features more compatible and robust.

[0035] (18); (19); (20); (twenty one); (twenty two); (twenty three); (twenty four).

[0036] (4) Finally, channel weighting factor is introduced. w Obtain the results of adaptive fusion: (25); in w The following is an expression: (26); in CRCS Represents convolution + Relu Activation function + convolution + Sigmoid Activation function.

[0037] Feature map visualization results are as follows Figure 4As shown, the features at different stages exhibit significant differences. The spatial domain features in the first row retain relatively clear boundary information, but due to interference from background information, the network produces false responses in the water area; for example, the grass texture around the second underwater animal is incorrectly activated. The frequency domain features in the second row have richer global semantics and can accurately locate the target's position, but the boundary details are weak, and the target outline is blurred, especially when color distortion occurs, making it impossible to accurately depict the boundary. The fused features generated by the frequency-spatial domain collaborative perception module, shown in the third row, combine global localization and local detail capabilities, significantly suppressing background false responses, restoring target edges and textures, and achieving accurate target localization in complex underwater scenes.

[0038] Cross-scale two-way interactive module: After processing by the frequency-spatial domain collaborative sensing module, the model can effectively suppress noise interference and enhance the structural representation of salient targets at both global and local levels. However, due to the scale diversity and structural differences of underwater targets, in order to further enhance the multi-scale representation capability of features and achieve full interaction of cross-layer information, this invention designs a cross-scale bidirectional interaction module, such as... Figure 5 As shown, this module consists of three parts: multi-scale convolutional feature extraction, cross-scale bidirectional interaction, and weighted fusion. It aims to achieve adaptive perception of targets at different scales through a bidirectional information flow mechanism. In this process, the fine-grained edge information of small-scale features and the semantic expression of large-scale features are effectively combined, enhancing the saliency response of targets at different scales.

[0039] Specifically, this module comprises three branches at different scales, each extracting features using convolutional kernels of different sizes. Small-scale kernels (3×3) capture detailed information; medium-scale kernels (5×5) are better suited for extracting structural features; and large-scale kernels (7×7) excel at capturing global information. The feature maps from these three scales are then concatenated to form a fused feature map. (27); (28); (29); in, FE This indicates the feature extraction module, which mainly includes operations such as convolutional layers, activation functions, and batch normalization. The subscript numbers indicate the size of the convolutional kernel.

[0040] To fully leverage the complementarity of features at different scales in terms of semantic representation and detail preservation, this invention proposes a bidirectional cross-scale interaction mechanism comprising channel interaction and spatial interaction. First, considering that large-scale features have a larger receptive field and stronger global semantic information, they are used to generate channel attention, thereby guiding the channel selection of small-scale features. This process is achieved through the following formula: This channel constraint can suppress noise inconsistent with global semantics at the detail level, while preserving edge and texture responses relevant to the target semantics. Second, given that small-scale features retain more refined boundary and positional information in the spatial dimension, they are used to generate spatial attention to correct the spatial response of large-scale features: This spatial guidance focuses global semantics on the target region, reducing semantic diffusion in the background, thereby improving localization accuracy. Finally, the enhanced interactive features are obtained: (30); (31); (32); This invention applies to small-medium ( ), medium-large ( ), small-large ( The three scales were interacted with to obtain three pairs of enhanced features. , and The final fused features are then convolved to obtain an enhanced output: (33)

[0041] Reverse region-boundary guided deformable attention module: Underwater imaging is commonly plagued by degradation factors such as color distortion and noise interference, often resulting in incomplete underwater target prediction regions and fragmented boundary responses during the decoding stage, severely limiting detection accuracy. To address this issue, this invention proposes a reverse region-boundary guided deformable attention module, aiming to enhance the model's structural awareness capabilities in underwater degradation scenarios from two complementary perspectives: region integrity and boundary refinement. This module consists of two parts: reverse region deformable attention and reverse boundary deformable attention. Its core lies not in simply superimposing region and boundary information, but in addressing the problem of highly unstable positive saliency and boundary prediction under underwater degradation conditions. It introduces a "reverse region-reverse boundary" as a more reliable structural constraint, guiding deformable convolution for dynamic sampling, thereby achieving collaborative modeling of region integrity and boundary refinement.

[0042] (1) Feature alignment: First, spatial alignment is performed on the features in the encoding and decoding stages. Specifically, the decoded features are adjusted to the same spatial resolution as the fused features in the encoding stage through upsampling operations, and then refined by concatenation, batch normalization and ReLU activation to obtain a unified fused feature representation.

[0043] (34); (35); in CBR This represents convolution, batch normalization, and ReLU activation functions.

[0044] (2) Reverse Region Deformable Attention: In underwater scenes, target regions often exhibit incomplete and unstable salient responses due to degradation effects, while background regions are more continuous and stable in spatial structure. Based on this observation, this invention proposes a reverse region deformable attention module. This module no longer relies on positive saliency regions as the main constraint, but explicitly models the background region through a reverse region mask. Specifically, a background mask image is first generated through thresholding. Then, the fused features are multiplied element-wise with the background mask image, and then a mask-aware deformable convolution is introduced under the guidance of the background mask. This convolution can adaptively adjust the sampling position, ensuring the continuity of the overall target structure and the distinguishability from the background.

[0045] (36); (37); (38); in This is the background mask image. w(p) Indicates the position of the convolution kernel p The weight, Indicates the center position of the convolution kernel. p This represents the position of a fixed sampling point relative to the center of the convolution kernel. This represents the offset generation function.

[0046] (3) Reverse Boundary Deformable Attention: To address the boundary ambiguity and discontinuity issues caused by underwater degradation, this invention further enhances contour modeling capabilities at the boundary level by proposing a reverse boundary deformable attention mechanism. Unlike traditional methods that directly enhance predicted boundaries, in underwater scenarios, the positive boundary response is often noisy and unreliable, and directly using it for constraints can easily amplify erroneous information. Therefore, we adopt a reverse boundary guidance mechanism, treating the boundary region as an unstable factor for explicit suppression, while utilizing non-boundary regions to provide a more stable structural reference.

[0047] First, the underwater saliency map is analyzed using the Sobel operator. Gradient calculations are performed to obtain edge responses in both the horizontal and vertical directions, and non-maximum suppression (NMS) is used to generate an initial boundary map. : (39); (40); (41); in for Sobel nuclear, NMS This indicates nonmaximum suppression.

[0048] Subsequently, to achieve reverse boundary guidance, the fused features will be... With reverse boundary Figure 1 - Element-wise multiplication is performed to suppress unstable boundary responses and obtain more reliable structural features. Building upon this, boundary-aware deformable convolution is introduced. By modulating the sampling offset through boundary responses, the convolution kernel adaptively shifts near the edges, enabling a dynamic sampling process for contour awareness. (42); (43); This design enables the convolution kernel to adaptively adjust the sampling position near the real target contour, which reduces the interference of noise boundaries on the sampling process and enhances the continuous modeling capability along the real contour direction, thereby effectively improving the accuracy and coherence of underwater target boundary details.

[0049] (4) Region-Boundary Collaborative Fusion: Finally, the output features of the RRDA and RBDA branches are fused in the decoding stage to generate refined saliency prediction results. Through collaborative optimization of regions and boundaries, the RRBDA module can achieve clear restoration and structural integrity reconstruction of salient target boundaries under complex underwater degradation conditions, providing a more robust and reliable feature representation for underwater saliency detection.

[0050] (44)

[0051] Table 1-1 lists the performance results of representative salient object detection methods on the USOD10K dataset. .

[0052] Table 1-2 lists the performance results of representative salient object detection methods on the USOD dataset. .

[0053] Quantitative Comparison: Tables 1-1 and 1-2 list the performance comparison results of representative salient object detection methods on the USOD10 and USOD datasets. Overall, the model proposed in this invention performs well in various evaluation metrics ( Fmean β , Fadp β , Emax ξ , Emean ξ , Eadp ξ , M Significant advantages were achieved in all aspects, fully verifying its robustness and accuracy in complex underwater environments. From the perspective of model size and computational efficiency, lightweight models (such as SATNet and RAMENet) have lower computational overhead, but their detection performance in degraded scenarios is generally insufficient; while heavyweight models (such as LAFB, SPDENet, TC-USOD, and UDNet) perform well in some metrics, but at the cost of a significantly increased number of parameters. In contrast, GeleNet, DPU, LBANet, and the model of this invention achieve a balance between performance and efficiency while maintaining appropriate parameters, with the model of this invention showing the best results. For example, on the USOD10K dataset, the model of this invention... Emax ξ , Emean ξ , Eadp ξ These represent improvements of 0.41%, 0.45%, and 0.5% respectively compared to DPU, with error indicators... M The performance was reduced by 4.12%; compared to LBANet, it was increased by 0.21%, 0.42%, and 0.72%, respectively. M The accuracy was reduced by 8.37%. These results demonstrate that the proposed frequency-space cooperative structure significantly improves detection accuracy without increasing model complexity. From the perspective of feature fusion and model design, some existing methods (such as DACNet, TC-USOD, and SPDENet) enhance feature representation through multimodal feature fusion (RGB and depth maps), but this easily introduces noise interference and feature redundancy, leading to decreased generalization. In contrast, this invention maintains a significant advantage on the USOD dataset: Emax ξ , Emean ξ , Eadp ξ These figures represent improvements of 0.67%, 0.92%, and 1.13% compared to LBANet, respectively. M It decreased by 13.43%; compared to UDNet, it increased by 0.50%, 0.59%, and 0.6%, respectively. M A decrease of 8.52%. It is worth noting that in some indicators (such as...) S αIn terms of detection performance, some models that incorporate explicit boundary supervision (such as UDCNet and LBANet) show a slight advantage, mainly due to the direct guidance of boundary constraints on detail learning. Future research could further explore the combination of boundary supervision mechanisms and frequency-space collaborative frameworks to improve boundary consistency and visual integrity. In summary, the quantitative results clearly demonstrate that simply stacking parameters or deepening the network structure is insufficient to effectively address the challenges posed by underwater degradation; while the structural design based on frequency-space collaborative modeling and efficient attention decoding can achieve a good balance between detection performance and computational efficiency, and significantly enhance the stability and generalization of the model in complex scenarios.

[0054] 2) Qualitative analysis: such as Figure 7 As shown, the qualitative comparison results of different methods on the USOD10K and USOD datasets are presented. Based on underwater degradation characteristics, the examples are divided into three typical scenarios: severe color distortion (rows 1–4), turbid water (rows 5–8), and complex underwater backgrounds (rows 9–12). In the severe color distortion scenario, illumination attenuation and color shift significantly reduce the contrast between the target and the background. Existing methods (such as DPU and LBANet) often suffer from incomplete saliency regions or blurred boundaries under these conditions, while this invention can generate saliency maps that are more consistent with the ground truth, maintaining better regional coherence. In the turbid water scenario, scattering effects and suspended particles cause blurred or even locally disappearing target boundaries, and most comparison methods are prone to boundary breaks or false background activation. Figure 7 As shown in lines 5–8, this invention can stably locate salient targets and maintain a clear overall structure even under highly turbid conditions, demonstrating the stronger robustness of the proposed model to turbidity degradation. In complex underwater background scenarios, existing methods generally suffer from inaccurate target localization or unclear outlines due to the high similarity between background texture and target appearance. In contrast, this invention can more accurately distinguish between the target and the background, and the generated saliency map is superior to the comparison methods in terms of boundary clarity and structural integrity. Overall, this invention demonstrates superior saliency detection performance compared to existing methods in three typical underwater degradation scenarios, verifying the generalization ability and robustness of the proposed model in complex underwater environments.

[0055] Table 2-1 Detection results of the model of this invention and other methods on the MAS3K dataset. .

[0056] Table 2-2 Detection results of the model of this invention and other methods on the RUWI dataset. .

[0057] Table 3-1 Detection results of the model of this invention and other methods on the RMAS dataset. .

[0058] Table 3-2 Detection results of the model of this invention and other methods on the SUIM dataset. .

[0059] 3) Cross-domain testing: To systematically verify the adaptability and generalization ability of the proposed invention in cross-domain underwater salient target detection, this study conducted a comprehensive experimental evaluation on four publicly available datasets with significant domain differences (MAS3K, RUWI, RMAS, and SUIM). The results are detailed in Tables 2 and 3. On the MAS3K dataset, where camouflaged targets account for a high proportion, the invention demonstrates significant advantages: S α The metric reached 0.8313, a 1.3% improvement over the LBANet method. and They increased by 2.19% and 1.72% respectively, while M A 10.65% reduction in performance. This superior performance validates the innovation of the frequency-spatial domain collaborative sensing module, which effectively decouples the spectral features of the target and background through frequency domain analysis, overcoming the inherent limitation of traditional spatial domain methods being susceptible to background interference in camouflaged scenarios. The more challenging RUWI dataset (real-world field scenes) further validates the robustness of this invention. Although the overall performance of various methods on this dataset is relatively low, this invention maintains a stable lead in several metrics. On the SUIM dataset, which is dominated by low-light scenes, the performance of this invention is relatively limited, a result that points the way for future research. Combining the experimental results of the four datasets, this invention demonstrates excellent cross-domain adaptability.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for underwater salient object detection based on frequency-spatial domain collaborative perception, characterized in that, It includes a frequency domain-spatial domain collaborative sensing module, a cross-scale bidirectional interaction module, and a deformable attention module guided by reverse region-boundary; The frequency-space domain collaborative sensing module includes an adaptive frequency sensing module and a frequency-space domain interactive fusion module; Adaptive frequency sensing module: extracts low-frequency and high-frequency components; low-frequency components are used to enhance global structural information and model the response of underwater target, while high-frequency components focus on enhancing edge details that are blurred due to turbidity and noise. Frequency-spatial domain interactive fusion module: It uses correlation guidance and bidirectional gating mechanism to allow frequency-spatial features to mutually correct and enhance each other during interaction. Spatial details are used to suppress noise uncertainty areas in the frequency domain response. The global structural prior provided by the frequency domain inversely constrains the perception range of spatial features, guiding them to focus on the underwater target area. The frequency-spatial domain interactive fusion module employs an attention mechanism, including channel attention and spatial attention. Channel attention adaptively adjusts the weights of different feature channels, highlighting channel features that contribute to saliency discrimination and suppressing invalid channels affected by color distortion or noise. Spatial attention enhances the response of salient regions while weakening complex background regions, helping the network to more accurately locate underwater targets. The expression is as follows: (10)； (11)； (12)； (13)； Where CA represents channel attention and SA represents spatial attention; Cross-modal correlation guidance is introduced during the fusion process. By calculating the similarity factor between spatial and frequency domain features, salient regions consistent across modalities are highlighted, and spurious salient regions in single modalities are effectively suppressed. The expression is as follows: (14)； (15)； (16)； (17)； C is the number of channels; the feature map after convolution is denoted as ; A bidirectional gating mechanism is designed, which modulates frequency domain features with spatial features to enhance their boundary and texture details; at the same time, frequency domain features are used to guide spatial features to improve global consistency, and the original modal features are introduced into the fusion result in the form of residuals, as shown in the following expression: (18)； (19)； (20)； (21)； (22)； (23)； (24)； Finally, a channel weighting factor w is introduced to obtain the adaptive fusion result: (25)； Where w represents the following: (26)； Where CRCS represents convolution + ReLU activation function + convolution + Sigmoid activation function; Cross-scale bidirectional interaction module: Fine-grained edge information of small-scale features is effectively combined with semantic expression of large-scale features to enhance the saliency response of targets at different scales; The reverse region-boundary guided deformable attention module introduces a "reverse region-reverse boundary" as a structural constraint to guide deformable convolution for dynamic sampling, thereby achieving collaborative modeling of region integrity and boundary refinement. The reverse region-boundary guided deformable attention module performs feature alignment, specifically by upsampling the decoded features... Adjust to integrate features with the coding stage Using the same spatial resolution, and after refinement processes such as stitching and convolution, batch normalization, and ReLU activation, uniform fused features are obtained. ; The reverse region-boundary guided deformable attention module includes a reverse region deformable attention module. This module no longer relies on positive saliency regions as the primary constraint, but instead explicitly models the background region through a reverse region mask. The specific steps are as follows: First, a background mask image is generated through thresholding. Secondly, the fused features are multiplied element-wise with the background mask image; Finally, guided by the background mask, a mask-aware deformable convolution is introduced. This convolution can adaptively adjust the sampling position, ensuring the continuity of the overall target structure and the distinction from the background. (36)； （37）； ； (38)； in As the background mask image, This represents the weight of the convolution kernel at position p. denoted by , where p represents the center position of the convolution kernel, and p represents the position of a fixed sampling point within the convolution kernel relative to the center. This represents the offset generation function. It is the offset vector of each convolution sampling point.

2. The method of claim 1, wherein, The adaptive frequency sensing module achieves adaptive decoupling between global structure preservation and local detail enhancement based on different underwater degradation modes. Specifically: The input image is processed by a convolutional layer to extract low-dimensional feature representations, and a high-level semantic guidance mechanism is introduced to multiply the high-level features with position awareness with the low-level features element by element. (1)； wherein represents an element-wise multiplication operation, represents adjusting the spatial dimension of the feature map to the spatial dimension of the feature map b ; the feature map after convolution is denoted as .

3. The method of claim 1, wherein, The adaptive frequency sensing module is configured with a dynamic threshold and frequency masking mechanism, specifically as follows: 1) Compressed contextual representations are obtained through global average pooling to preserve semantic information and perceive overall scene characteristics; 2) Construct a lightweight prediction module consisting of two convolutional layers and nonlinear activation to perform nonlinear mapping on features; 3) Introduction Sigmoid Activate the generation of two normalized scaling coefficients and It is used to adaptively adjust the polar coordinate radius threshold, thereby dynamically dividing low-frequency and high-frequency components; (2)； (3)； in Cov , GAP and These represent convolutional layers, global average pooling, and... Sigmoid Activation function R low H represents the low-frequency radius, and H is the image height. 4) After obtaining the low-frequency radius, two indicator function masks were constructed based on the polar coordinate grid to distinguish between low-frequency and high-frequency components. The indicator functions for the low-frequency and high-frequency regions are defined as follows: (4)； (5)； (6)； in, x grid and y grid Let I represent the normalized grid coordinates in the horizontal and vertical directions, respectively, with a value range of [-1, 1]. A ( x ) represents a set A The indicator function on, if and only if x ∈ A The value is 1 at time. 5) After obtaining the high-frequency and low-frequency masks, process the feature map. Frequency domain features are obtained by performing a fast Fourier transform. Then, based on the low-frequency mask and the high-frequency mask, Enhancement is performed, and finally, the low-frequency and high-frequency components are transformed back from the frequency domain to the spatial domain using inverse Fourier transform to obtain separated low-frequency feature maps. and high-frequency feature maps ; 6) Introduce hyperparameters to control the relative proportion of low and high frequency components to get the final output : (9)。 4. The method of claim 1, wherein, The cross-scale bidirectional interaction module includes multi-scale convolutional feature extraction, cross-scale bidirectional interaction, and weighted fusion; Multi-scale convolutional feature extraction includes three branches of different scales. Each branch extracts features through convolution kernels of different sizes. The feature maps of each branch are stitched together to form a fused feature map. Cross-scale bidirectional interaction includes two parts: channel interaction and spatial interaction. Large-scale features are used to generate channel attention, which guides the channel selection of small-scale features. Small-scale features are used to generate spatial attention to correct the spatial response of large-scale features, as shown in the following expression: (30)； (31)； (32)； For small-medium ( , ), medium-large ( , ), small-large ( , The three scales were interacted with to obtain three pairs of enhanced features. , and ; The final fused features are then convolved to obtain an enhanced output: (33)。 5. The method of claim 1, wherein, The reverse-region-boundary guided deformable attention module includes reverse-boundary deformable attention, which treats boundary regions as unstable factors for explicit suppression, while utilizing non-boundary regions to provide a more stable structural reference. The steps are as follows: 1) on the underwater saliency map Gradient computation, edge response in horizontal and vertical directions, and non-maximum suppression to generate the initial boundary map : 2) To implement reverse boundary guidance, fuse features with reverse boundary map element-wise multiplication to get reliable structure features ; 3) Introducing boundary-aware deformable convolution, which modulates the sampling offset through boundary response, enabling the convolution kernel to generate adaptive displacement near the edge, thus realizing a dynamic sampling process for contour awareness: (42)； ； (43)。 6. The method of claim 5, wherein, The reverse region-boundary guided deformable attention module includes a region-boundary collaborative fusion module, which fuses the reverse region deformable attention output with the reverse boundary deformable attention output to generate refined saliency prediction results, as shown in the following expression: (44)。

Images