GME: A Multigranularity Attention and Mamba-Driven Approach for Underwater Image Enhancement

Abstract

In complex underwater environments, the selective absorption of light by the water medium severely impairs scene context and color fidelity, leading to distorted information and imbalanced colors in the acquired underwater images. Such degradations pose substantial challenges to underwater image enhancement (UIE). In this article, we propose a multiscale U-shaped network termed GME to achieve more effective UIE. Specifically, GME includes three essential components: 1) an enhancement layer (EL) with a group-mix attention (GMA) module that characterizes short-range local textures and medium-range regional context through grouped feature partitioning and mixed multigranularity matching to capture local and regional dependencies at multiple granularities for fine-detail recovery; 2) an enhancement block (EB) consists of a visual state-space (VSS) module and a squeeze-and-excitation (SE) unit which constructs input-aware state transitions to encode global contextual information and learns channel attention weights to reweight the feature channels, thus modeling the long-range spatial interactions and recalibrating the channel responses effectively; and 3) an adaptive enhancement module (AEM) that learns region-aware modulation coefficients to differentiate flat areas, edges, or complex regions modulates the residual enhancement strength adaptively to suppress over-enhancement artifacts in the challenging regions. Extensive experiments on multiple underwater image benchmarks show that GME achieves superior quantitative performance over state-of-the-art UIE methods and delivers visually more natural contrast and color appearance. The source code is available at https://github.com/OUC-AI/GME