Underwater image restoration is significant for various applications such as ecological evaluation, exploration, searching and rescue operations, and autonomous vehicle navigation. In underwater environments, spatial images are frequently degraded as a result of light scattering, absorption, sensor noise, and reduced contrast. This research proposes a whole framework with deep learning that simultaneously performs restoration and enhancement. At a single stage, solving the problems of underwater image degradation in a holistic approach. The core of the proposed approach in this study is a Denoising Convolutional Neural Network (DnCNN) architecture. Where the suppression of noise takes place with extreme focus on significant detail by an advanced non-local attention mechanism. For the further natural color restoration, multi-color space transformations RGB, LAB, and HSV. Which come into play for enhancing the contrast adjustment and color correction of contrast and a correction of colors enabling effective correction of deep-sea views. The framework takes advantage of both synthetically modified images and actual underwater images for model training which offers enhanced generalization for various settings. For the evaluation of the developed approach, two datasets of underwater images, EUVP and LSUI, were used. For the evaluation of the developed approach, two datasets of underwater images, EUVP and LSUI, were used. For the EUVP dataset, this model produces a PSNR of 30.77 dB, an SSIM of 0.892, RMSE of 0.065, and NIQE of 3.52. It produces a PSNR of 29.90 dB, SSIM of 0.881, RMSE of 0.071, and NIQE of 3.82 for the LSUI dataset. As mentioned earlier these results outperform the baseline DnCNN model and show consistent performance with varying underwater conditions. Accompanying the performance results, the model achieves high fidelity results while being lightweight, real-time processing.
Ali et al. (Tue,) studied this question.