Two-Stage Progressive Underwater Image Enhancement
Junjun Wu, Xilin Liu, Ningwei Qin, Qinghua Lu, Xiaoman Zhu
Abstract
Underwater image enhancement (UIE) is a challenging problem involving various aspects of image degradation, such as color scattering, low contrast, and haziness. In this study, we present a method named two-stage progressive enhancement network (TPENet) to address these issues. We outline the challenges faced in UIE and introduce how TPENet tackles them. TPENet adopts a two-stage network architecture that combines the extensive context learning capabilities of encoders–decoders and the spatial-detail preservation capabilities of the original resolution network. In the first stage, we design a densely fused encoder–decoder subnetwork that focuses on addressing color distortion and low contrast issues. In the second stage, we introduce an original resolution subnetwork (ORSNet) and tackle the haziness problem in underwater images through an image dehazing auxiliary task. To highlight local features and pass them for further enhancement in the next stage, we also introduce a multicolor space supervised attention module. Through extensive experimental results, we validate the outstanding performance, generalization ability, and positive impact on other visual tasks of the proposed method. Additionally, we conduct ablation studies to demonstrate the contributions of key components.