A Hybrid Learning Framework for Underwater Image Enhancement

Underwater imaging facilitates the exploration of the underwater environment. However, irregular optical absorption and light scattering in water, ranging from clear to highly turbid conditions, often result in low visibility, color distortion, and blurriness in underwater images (UWIs). Conventional UWI enhancement methods are limited by inefficient physical modeling, while deep learning-based approaches are constrained by the scarcity of paired training datasets. In this work, we propose a hybrid learning framework for UWI enhancement that leverages the usefulness of both conventional and deep learning-based techniques. At first, we preprocess the UWIs using a revised underwater physical model, contrast adaptive histogram equalization (CLAHE), gamma correction (GC), and single-scale retinex (SSR) techniques. These preprocessed images, together with the corresponding raw UWIs, are then used to train a convolutional neural network (CNN) to generate confidence maps for each enhancement result. Simultaneously, we employ a second CNN-based feature transformation network (FTN), which consists of four feature transformation units (FTU). We train each FTU using a raw UWI and its corresponding preprocessed image to obtain confidence maps. In order to exploit the complementary strengths of the revised underwater physical model confidence and refined maps (from CNN and FTN), we use them to scale the remaining information maps. Finally, we perform additive fusion of the scaled maps to obtain enhanced UWI. Experiments on diverse UWI datasets, including the real-world image enhancement (RIE) dataset (which contains a variety of degradation effects with greenish, bluish, and green-bluish tones), the underwater image enhancement benchmark (UIEB), and the enhancement of underwater visual perception (EUVP) dataset, demonstrate the enhanced generality and effectiveness of the proposed approach.