UP-GAN: Channel-spatial attention-based progressive generative adversarial network for underwater image enhancement

Abstract

Focusing on severe color deviation, low brightness, and mixed noise caused by inherent scattering and light attenuation effects within underwater environments, an underwater-attention progressive generative adversarial network (UP-GAN) is innovated for underwater image enhancement (UIE). Salient contributions are as follows: (1) By elaborately devising an underwater background light estimation module via an underwater imaging model, the degradation mechanism can be sufficiently integrated to fuse prior information, which in turn saves computational burden on subsequent enhancement; (2) to suppress mixed noise and enhance foreground, simultaneously, an underwater dual-attention module is created to fertilize skip connection from channel and spatial aspects, thereby getting rid of noise amplification within the UIE; and (3) by systematically combining with spatial consistency, exposure control, color constancy, color relative dispersion losses, the entire UP-GAN framework is skillfully optimized by taking into account multidegradation factors. Comprehensive experiments conducted on the UIEB data set demonstrate the effectiveness and superiority of the proposed UP-GAN in terms of both subjective and objective aspects.