CURE-Net: A Cascaded Deep Network for Underwater Image Enhancement
Xiaowen Cai, Nanfeng Jiang, Weiling Chen, Jinsong Hu, Tiesong Zhao
Abstract
Underwater imaging is usually degraded by low contrast or color cast due to light absorption and scattering. This fact limits the accuracy of object recognition in underwater and marine environments. In this article, to address this issue, we propose a CURE-Net that progressively improves degraded underwater images in a coarse-to-fine way. To be specific, our CURE-Net consists of three cascaded subnetworks. The first two subnetworks use attention and gate fusion to learn multiscale contextual information, whereas the third subnetwork preserves fine spatial details. A detail enhancement block and a supervised recovery block are deployed between subnetworks to further improve the recovery of colors and details. Our proposed CURE-Net is extensively evaluated on popular data sets and achieves state-of-the-art performance. Compared with its peers, our method is also more friendly to object recognition—underwater objects are more easily recognized after being processed by our method.