Unsupervised Deep-Learning Approach for Underwater Image Enhancement

Alejandro Rico Espinosa, Declan George David McIntosh, Alexandra Branzan Albu

Abstract: Underwater images often contain color casting and blurriness which reduce the quality. State-of-the-art shows different deep-learning models to handle these degradations. However, they required ground truth to train, which is impossible to acquire when studying underwater images. We present an unsupervised deep-learning approach for underwater image enhancement based on the mathematical model of a hazy image. This allows us to train networks without the need for a reference image. We use three networks to estimate the transmission map, the atmospheric light, and the enhanced image and propose a compound loss function to train our approach accurately. We achieve state-of-the-art results in the structural similarity index (SSIM) while performing optimally nearly real-time inference speeds.