Enhanced visual perception for underwater images based on multistage generative adversarial network

Underwater images often suffer from color distortion and low contrast, which dramatically affects the target detection and measurement tasks in the underwater context. In this paper, we present a multistage generative adversarial network for better visual perception of underwater images. Extensive multi-scale context feature learning and high-precision restoration of spatial details are implemented stage by stage. Rich context features are learned based on the encoder and decoder architecture. Spatial details are restored through a pixel restoration module based on original images. Through channel attention module used between multistages, cross-stage feature utilization is realized. More notably, we introduce Gaussian noise into the generator, which enriches the details of images, and the relative discriminator, which promotes the generated image to have more realistic edges and textures. Experimental results demonstrate the superiority of our method over state-of-the-art methods in terms of both quantitative metrics and visual quality. In particular, we applied our method to natural underwater scenes. The results confirm that our method can effectively improve the efficiency of downstream tasks.