Real-time underwater image resolution enhancement using super-resolution with deep convolutional neural networks

In this paper, a two-step image enhancement is presented. In the first step, color correction and underwater image quality enhancement are conducted if there are artifacts such as darkening, hazing and fogging. In the second step, the image resolution optimized in the previous step is enhanced using the convolutional neural network (CNN) with deep learning capability. The main reason behind the adoption of this two-step technique, which includes image quality enhancement and super-resolution, is the need for a robust strategy to visually improve underwater images at different depths and under diverse artifact conditions. The effectiveness and robustness of the real-time algorithm are satisfactory for various underwater images under different conditions, and several experiments have been undertaken for the two datasets of images. In both stages and for each of image datasets, the mean square error (MSE), peak signal to noise ratio (PSNR), and structural similarity (SSIM) evaluation measures were fulfilled. In addition, the low computational complexity and suitable outputs were obtained for different artifacts that represented divergent depths of water to achieve a real-time system. The super-resolution in the proposed structure for medium layers can offer a proper response. For this reason, time is also one of the major factors reported in the research. Applying this model to underwater imagery systems will yield more accurate and detailed information.