AquaIE: A Lightweight Framework for Underwater Image Enhancement Combining Mamba Blocks

Underwater images often suffer from color distortion, low contrast, and blurring due to light absorption and scattering in water. This paper presents AquaIE, a lightweight generative adversarial network (GAN) framework that combines novel Mamba state-space modules with an encoder–decoder architecture to address these challenges. The proposed AquaIE integrates Mamba blocks into a U-Net backbone to capture long-range dependencies and global context, enhancing color fidelity and structural preservation in degraded underwater scenes. In addition, a spiral adversarial training strategy is introduced, wherein the generator’s output is iteratively refined over multiple cycles, effectively acting as an implicit data augmentation and improving robustness to domain shifts. A comprehensive loss function, including adversarial, structural similarity, angular color consistency, and perceptual components, guides the network to produce realistic and high-quality enhancements without introducing artifacts or color distortion. Extensive experiments on multiple underwater image benchmarks demonstrate that AquaIE achieves state-of-the-art performance in terms of contrast improvement, color correction, and detail recovery. For instance, on the challenging EUVP Underwater Dark dataset, AquaIE boosts the underwater image quality measure (UIQM) and colorfulness (UICM) metrics by a significant margin over existing methods. Qualitative results further showcase AquaIE’s ability to restore vivid colors and sharp details in diverse underwater conditions. The proposed AquaIE framework is efficient and generalizes well to real-world underwater images, making it a promising solution for underwater image enhancement in practical applications.