UGIT-CLRNet: Hybrid Transformer-CNN framework for Underwater Image Enhancement

Abstract

Light absorption and scattering cause severe degradation of underwater images and colour distortion, low contrast and loss of structural information. Traditional ways of improving images using histogram equalisation create artefacts and do not rectify global illumination aberrations. In the present paper, we suggest UGIT-CLRNet, a hybrid deep learning model with one Vision Transformer (ViT) branch to capture global contextual information and one branch based on the Convolutional Neural Network (CNN) to refine the clarity of the information. End-to-End joint optimization is implemented in the architecture to both restore colour balance and sharpen fine textures. On the EUVP paired dataset, the PSNR and SSIM of UGIT-CLRNet are 29.53 dB and 0.9336 respectively, which is far beyond the classical and CNN-based baselines. Non-reference metrics (UCIQE and UIQM) also prove the perceptual quality improvement.