Degradation-Driven Underwater Image Enhancement

Abstract

The use of robotic equipment in modern underwater activities is an usual practice. Maintenance, robotic inspection and environment and biological research are some examples of applications and the acquisition of the images and videos is a common procedure. Frequently, these images suffer with light attenuation and turbidity of the water requiring enhancement or restoration. In this work, we present a method for underwater image enhancement based on deep learning and inspired in the Underwater Image Formation Model but color space-contextualized. The strategy misleads a neural network, generating a fake and distorted output image that replaces the real network output in the loss function. The algorithm does not estimate physical parameters, but explores similar representations in the color space. The only information required is the input image and the methodology require no ground-truth and unsupervised learning is adopted. Only real underwater images are used and the results indicate the effectiveness of the method in color preservation, sharpness and contrast improvement.