Underwater Image Enhancement Using Dual Adversarial Network

Abstract

Since the degradation of the image has seriously constrained the development of marine research, the underwater image enhancement has been paid more and more attentions. Due to the diversity of underwater images (for example, underwater images show different attenuation and color bias in different scenes) and the lack of underwater datasets, most existing methods usually show satisfactory results on some kinds of underwater types. To solve the problems, we built a novel model, including two adversarial network blocks, to learn the essential content features of multiple underwater types and restore high quality images. We trained the model under the synthetic dataset based on Jerlov underwater type image dataset. Experimental results show that the model not only outperforms most previous methods in PSNR and UIQM but also shows the generalization ability.