Excavate the Potential of Single-Scale Features: A Decomposition Network for Water-Related Optical Image Enhancement

Underwater image enhancement techniques typically rely on multiscale feature extraction to restore images degraded by light absorption and scattering. This article challenges that dominant paradigm by demonstrating that a meticulously designed single-scale architecture can achieve highly comparable performance to multiscale counterparts, while significantly reducing model complexity. We propose the single-scale decomposition network (SSD-Net), an innovative framework that explores the full potential of single-scale representations. SSD-Net introduces an asymmetric pipeline to decouple the input into a scene-intrinsic clean layer and a medium-induced degradation layer. This is achieved through two core synergistic modules: first, the parallel feature decomposition block, which utilizes a sparse Transformer and CNNs for dual-branch feature disentanglement, and second, the bidirectional feature communication block, which enables cross-layer residual interactions for mutual refinement. This design preserves decomposition independence while establishing dynamic information pathways, maximizing the efficacy of single-scale features. Compared to state-of-the-art multiscale approaches, SSD-Net achieves superior enhancement quality with substantially fewer parameters and computations.