Deep Learning Based Underwater Image Enhancement Using Hybrid CNN GAN Approach with Self Attention Mechanism

Abstract

Marine exploration, environmental monitoring as well as autonomous underwater vehicles need underwater imaging, however, it is severely degraded by light absorption, scattering and color distortion. Histogram equalization and contrast stretching are traditional enhancement techniques that do not generalize over different underwater conditions. This paper is implemented as a deep learning approach that utilizes Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) to solve the problem of improving underwater images with high precision. It trains the proposed model on a highly diverse set of underwater images and uses an adaptive loss function that helps preserve structural integrity and natural colors while removing noise. The system essentially restores visibility and contrast of degraded underwater scenes through attention mechanisms along with transfer learning integration. We compare state of the art in terms of PSNR, SSIM and Underwater Color Image Quality Evaluation (UCIQE) score, achieving superior performance. It is thus suitable for real time processing application in practical use such as underwater robotics and marine conservation. The future work will be on expanding the dataset and improving the model robustness with respect to different underwater conditions.