Interpretable Underwater Image Enhancement based on Convolutional Neural Network

Abstract

An underwater image suffers from degradation due to the physical attributes of water. The enhancement of degraded underwater images is an important area of research. Several researchers have been using machine learning-based models for enhancement. But, the network models are solely based on training data and the results are difficult to explain. Here, we present a novel enhancement technique for underwater image utilizing a set of enhancement functions and a Convolutional neural network(CNN). The four functions are blended to create the resultant enhancement function. The proposed network is interpretable in the sense that the work of the four functions are easily understandable and they can efficiently enhance different part of an underwater image. The CNNs are used to tune the parameters of the functions depending on the training data. The performance of the proposed method is quite efficient compared to the recently published methods on standard dataset.