An End-to-End Underwater-Image-Enhancement Framework Based on Fractional Integral Retinex and Unsupervised Autoencoder

As an essential low-level computer vision task for remotely operated underwater robots and unmanned vehicles to detect understand the environment, image enhancement is facing challenges of light scattering, absorption, distortion. Instead using a specific imaging model mitigate degradation images, we propose end-to-end underwater-image-enhancement framework that combines fractional integral-based Retinex encoder–decoder network. The proposed variant aims alleviate haze color distortion in input while preserving edges large extent by utilizing modified integral filter. network with channel-wise attention modules trained unsupervised manner overcome lack paired datasets designed refine output Retinex. Our was evaluated under qualitative quantitative metrics on several public yielded satisfactory results evaluation set.