Technical Summary: Learning Dynamic Structural Specialization for Underwater Salient Object Detection

1. Problem Statement

Underwater Salient Object Detection (USOD) is a fundamental task for underwater visual scene understanding and vision-guided robotics. However, existing methods face significant challenges due to severe image degradations inherent to underwater environments, such as light absorption, scattering, color distortion, low contrast, and noise. These degradations weaken visual cues, leading to inaccurate object localization, fragmented salient regions, and coarse boundary predictions.

The paper identifies a specific limitation in current RGB-based USOD methods: boundary–region entanglement. Most existing approaches process boundary-sensitive (high-frequency) and region-coherent (low-frequency/contextual) structural cues within a unified, homogeneous representation. This shared processing causes these complementary but distinct representational demands to interfere with one another. In degraded underwater scenes, where boundary cues are weak and regional consistency is disrupted, this entanglement results in fragmented salient regions and blurred boundaries. While some methods introduce depth information (RGB-D) to compensate, this increases system complexity and relies on unreliable depth estimation or pseudo-depth generation.

2. Methodology: DSS-USOD

The authors propose DSS-USOD, a novel RGB-based USOD framework built upon dynamic structural specialization. The core philosophy is to explicitly disentangle boundary-sensitive and region-coherent representations and dynamically coordinate their contributions based on local structural context, rather than relying on auxiliary modalities like depth.

The architecture consists of three main components:

A. Multi-scale Representation Learning

The model employs a hierarchical encoder (based on PVTv2) to extract multi-scale feature maps from the input underwater image. These features are projected into a unified channel dimension and aggregated via a top-down feature pyramid to form a shared base representation ($F_b$).

B. Dynamic Structural Specialization

Instead of decoding $F_b$ directly, the framework decomposes it into two complementary branches:

1. Boundary-Sensitive (BS) Branch: Designed to capture high-frequency boundary details. It processes $F_b$ through two parallel paths: a fixed Laplacian operator (providing an explicit inductive bias for boundaries) and a learnable high-pass transformation (adapting to task-specific patterns). This enhances structural responses around object boundaries.
2. Region-Coherent (RC) Branch: Designed to capture low-frequency structural consistency and region-level coherence. It utilizes a residual dual-scale anisotropic large-kernel modeling strategy (using factorized anisotropic depthwise operators with kernel sizes 7 and 15) to aggregate contextual information over extended spatial ranges. This helps suppress isolated noise and ensures complete salient regions.

Spatial Coordination Module (SCM):
Recognizing that different spatial locations require different structural preferences (e.g., boundaries need high-frequency cues, while interiors need region coherence), the SCM adaptively regulates the contributions of the BS and RC branches.

• It computes a branch discrepancy map ($D$) by measuring the element-wise difference between the BS and RC features.
• It integrates $F_b$, the predicted boundary response, and $D$ to generate a spatial weight map ($w$).
• The final coordinated representation ($F_d$) is computed as a weighted sum: $F_d = w \otimes F_{bs} + (1-w) \otimes F_{rc}$.

C. Progressive Prediction Refinement (PPR)

The model recovers the full-resolution saliency map in a coarse-to-fine manner:

1. Predicts low-resolution logits from $F_d$.
2. Generates a coarse full-resolution prediction by combining $F_d$ and the low-resolution logits.
3. Performs residual refinement by exploiting RGB appearance and shallow spatial features to correct local inaccuracies near boundaries and ambiguous transitions.

D. Cooperative Structural Supervision

To stabilize the learning of specialized branches and the spatial coordination, a cooperative supervision strategy is employed during training:

• Multi-level Saliency Loss: Applied to final, coarse, and low-resolution predictions.
• Boundary Loss: Supervises the BS branch using ground-truth boundary maps.
• Coordination Loss: Supervises the SCM by generating a target map that highlights boundary/transition regions (where BS should dominate) and homogeneous interiors (where RC should dominate).

3. Key Contributions

1. Identification of Boundary–Region Entanglement: The paper characterizes the insufficient differentiation between boundary-sensitive and region-coherent demands in RGB-based USOD as a key challenge, proposing that explicit disentanglement is necessary for robust performance.
2. Dynamic Structural Specialization Framework: The authors propose DSS-USOD, which decomposes shared features into specialized BS and RC branches and uses a Spatial Coordination Module to adaptively balance them based on local context.
3. Cooperative Structural Supervision: A training strategy is introduced to promote effective branch specialization and stabilize the spatial coordination mechanism.
4. Performance and Practicality: The method achieves state-of-the-art results on benchmark datasets using only RGB input, avoiding the computational overhead and unreliability of depth sensors. It is also validated on a real-world underwater robot.

4. Experimental Results

The authors evaluated DSS-USOD on two benchmark datasets: USOD10K and USOD.

• Quantitative Performance: DSS-USOD (specifically the DSS-USOD-S variant) achieved superior performance across four metrics (S-measure, E-measure, maxF, and MAE) compared to 40 state-of-the-art methods, including RGB-based, RGB-D-based, and diffusion-based approaches.
  • On USOD10K, it achieved an S-measure of 0.9338 and MAE of 0.0165.
  • Notably, it outperformed strong RGB-D methods (e.g., FSCDiff, SearchNet) while relying solely on RGB input.
• Generalization: When tested on six terrestrial SOD benchmarks (e.g., DUTS-TE, ECSSD) without fine-tuning, DSS-USOD demonstrated competitive cross-domain performance, particularly in minimizing pixel-level errors (MAE), suggesting the learned structural cues are transferable.
• Ablation Studies:
  • Removing either the BS or RC branch degraded performance, confirming their complementary nature.
  • Replacing the SCM with fixed averaging or global weighting resulted in lower performance, validating the need for spatially adaptive coordination.
  • Removing cooperative supervision led to less accurate boundary and region predictions.
  • The Progressive Prediction Refinement (PPR) strategy was shown to be critical for achieving high-resolution accuracy.
• Real-World Deployment: The model (DSS-USOD-T) was deployed on an underwater robot equipped with an NVIDIA Jetson TX2 NX. It achieved an inference speed of 21 FPS, successfully performing real-time visual target inspection and highlighting foreground objects while suppressing underwater noise.

5. Significance and Claims

The paper claims that DSS-USOD offers a new perspective on RGB-based USOD by demonstrating that explicitly disentangling and dynamically coordinating complementary structural representations is more effective than unified modeling or relying on auxiliary depth modalities.

The authors assert that their approach addresses the fundamental issue of boundary–region entanglement, enabling the model to balance boundary precision and region coherence even under severe underwater degradation. The successful real-world deployment on a resource-constrained robot validates the practical effectiveness of the method for vision-guided underwater tasks, such as environmental monitoring and infrastructure inspection, without the need for complex depth sensing hardware.