MSA-CycleGAN: Multiscale structure-aware CycleGAN for color cast removal in underwater images

Technologies for underwater image enhancement are vital to marine resource development, robotic exploration and ecological protection. Due to light absorption and scattering in water, underwater images often suffer from low contrast, blurred details and color deviations. Though many enhancement methods exist, most fail to adequately consider the importance of multi-scale analysis of images, thereby restricting their ability to capture diverse feature information in real-world underwater environments. This paper introduces a bidirectional cycle generative adversarial network (CycleGAN) model named MSA-CycleGAN, designed to address these challenges. It effectively restores image clarity, corrects color deviations, and mitigates distortions caused by underwater environments, while suppressing degradation features and preserving intricate detail information. This is achieved by incorporating multi-scale Laplacian pyramid decomposition, channel attention mechanism (CAM), and an image dehazing module into the multiscale structure-aware generator and the detail-structure aware discriminator, together with the application of bidirectional cycle loss and collaborative optimization of multiple loss functions. Experiments and practical case studies show our method outperforms mainstream underwater enhancement approaches in qualitative and quantitative metrics. Additional underwater object detection tests further verify that MSA-CycleGAN achieves higher detection accuracy with faster inference and a lightweight model.