Underwater Image Processing with Normalized AttUNet

The goal is enhancement of images with different water types and to achieve visual quality improvement of the underwater images. Poor visibility in underwater images is a result of color casts, low contrast, and particle dispersion, all of which are caused by wavelength-dependent and distance-dependent attenuation. To address these degradation issues and to learn water type desensitized features, normalization is applied along both spatial and channel dimensions. In the spatial dimension, instance whitening is applied, to decorrelate activations. Moreover, channel-wise normalization is used to manipulate the first two moments of the channel activations. Then, attention is used to provide weights to certain features based on their importance. We proposed a lightweight model named Normalised AttUNet to provide fast processing time for diverse water type images, while still maintaining comparable PSNR.