Advancing Underwater Image Quality Enhancement through Hybrid Deep Learning Architectures

Abstract

Image quality enhancement in underwater settings is important for marine sciences, environmental monitoring, and the use of autonomous underwater vehicles. Lighting issues such as scattering, absorption and turbidity can have adverse impact on underwater imagery; these factors reduce the clarity of the picture, distort colors, and increase noise levels. Traditional enhancement methods often find it difficult to resolve these issues. We present a hybrid deep learning approach that fuses CNN and transformer based models to address the challenge of improving the quality of underwater images. In our approach, CNNs extract features and enhance spatial quality, while accumulation of long range and global contextual information is done at the transformer stage. The proposed model is trained on a set of diverse underwater images with robust performance achieved due to supervised and unsupervised learning approaches. Experimental data shows that our method surpasses the best currently available methods in clarity, color and shape restoration as well as PSNR, SSIM and UCIQE scores. The hybrid system greatly reduces the effects of the degradation of underwater images and increases the visibility and recognition of underwater objects. We open a new direction in underwater image processing with deep learning by advancing the mark for models that balance local perceptual and global semantic cues. In the future we will look into moving towards real time system as well as domain adaptation more broadly.