NovaNet-SSL: A Self-Supervised DDPM Model for Underwater Image Enhancement

Marine images suffer from severe tint of color and blurry feature caused by distortion in water. Enhancement of such images is necessary to be developed for the improvement of quality vision tasks underwater. This paper proposes a self-supervised diffusion model, called NovaNet, for real-time enhancement of underwater images (UIE). This technique uses the power of DDPM and optimized the process using self-supervised learning. NovaNet accelerates the inference process without a compromise on enhancement quality; it performs faster and more accurately than any other generative model. Unlike conventional DDPM models, which suffer from slow processing and color bias while used for UIE, this approach integrates self-supervised learning to effectively guide the model, improving the quality of image restoration. The novelty in this approach is self-supervised learning with diffusion-based image restoration to produce results of high quality, that is, enhanced quality output and fast processing speed. NovaNet demonstrates a 20.34% improvement in PSNR, a 0.5351% gain in SSIM, and achieves a 0.036% faster processing speed compared to existing models. This proposed model is tested against several challenging real-world underwater datasets, which are characterized to have superior performance regarding enhanced quality and real-time process capabilities.