A Review on Underwater Image Enhancement Using Deep Residual Framework

There are various factors such as absorption, refraction & the phenomenon of scattering of light by particles suspended in water that are responsible for distorted colors, low contrast & blurred details of original underwater images. The traditional approaches include pre-processing the image using a descattering algorithm. The super-resolution (SR) method is applied. But this method has limitation that major part of the high frequency information is lost during descattering. This paper comes up with a solution for underwater image enhancement using a deep residual framework. Firstly, the generation of synthetic underwater images takes place for which cycle-consistent adversarial networks (CycleGAN) is employed. Further, these synthetic underwater images are used as training data for convolution neural network models. Secondly, the introduction of very-deep super-resolution reconstruction model to underwater resolution applications is carried out. Using this, the underwater Resnet model is proposed. It acts as a residual learning model for underwater image enhancement operations. Furthermore, the training mode & loss function are improved. Then, a multi-term loss function is formed which comprises of proposed edge difference loss & mean squared error loss. An asynchronous training mode is also being proposed that improves the performance of the multi-term loss function. Lastly, the discussion of the impact of batch normalization takes place. After a comparative analysis & underwater image enhancements, we can say that detailed enhancement performance & color correction of these proposed methods are much efficient & superior to that of previous traditional methods & deep learning models.