CFPNet: Complementary Feature Perception Network for Underwater Image Enhancement

Images shot underwater are usually characterized by global nonuniform information loss due to selective light absorption and scattering, resulting in various degradation problems, such as color distortion and low visibility. Recently, deep learning has drawn much attention in the field of underwater image enhancement (UIE) for its powerful performance. However, most deep learning-based UIE models rely on either pure convolutional neural network (CNN) or pure transformer, which makes it challenging to enhance images while maintaining local representations and global features simultaneously. In this article, we propose a novel complementary feature perception network (CFPNet), which embeds the transformer into the classical CNN-based UNet3+. The core idea is to fuse the advantages of CNN and transformer to obtain satisfactory high-quality underwater images that can naturally perceive local and global features. CFPNet employs a novel dual encoder structure of the CNN and transformer in parallel, while the decoder is composed of one trunk decoder and two auxiliary decoders. First, we propose the regionalized two-stage vision transformer that can progressively eliminate the variable levels of degradation in a coarse-to-fine manner. Second, we design the full-scale feature fusion module to explore sufficient information by merging the multiscale features. In addition, we propose an auxiliary feature guided learning strategy that utilizes reflectance and shading maps to guide the generation of the final results. The advantage of this strategy is to avoid repetitive and ineffective learning of the model, and to accomplish color correction and deblurring tasks more efficiently. Experiments demonstrate that our CFPNet can obtain high-quality underwater images and show superior performance compared to the state-of-the-art UIE methods qualitatively and quantitatively.