CA-Net: Cascaded Adaptive Network for Underwater Image Enhancement

Abstract

Due to light absorption and scattering, underwater images often suffer from low contrast, blurry details, and color deviation. Various enhancement methods have been developed, but many fail to improve image quality effectively and sometimes create unnatural effects. To tackle such a problem, we propose a novel method, namely the Cascaded Adaptive Network (i.e., CA-Net), to comprehensively enhance the quality of underwater images. Specifically, our network adopts a cascaded enhancement architecture consisting of three stages (coarse feature restoration, feature aggregation, and color refinement). First, we use a detail restoration (DR) module and channel balance module to recover spatial details and correct color distortion, respectively, in the first stage. Particularly, the detail guidance unit of DR employs encoder features to steer the decoder features to focus more on the spatial details of objects. Second, to promote the fusion of fine details and color features, we deploy a context attention (CA) module and an adaptive feature fusion (AFF) module in the stage of feature aggregation. CA extracts detailed restoration features and long-range dependencies in images, guiding the fusion process in the subsequent AFF. Lastly, to guarantee natural colors, we use a global color rendering module in the stage of color refinement, which adaptively groups and tunes the image channels. Experiments on public data sets show that CA-Net significantly outperforms existing methods, making it highly effective for underwater image enhancement.