MLFFNet: Multi-level Feature Fusion Net for Underwater Image Enhancement

Abstract

Underwater image enhancement has been attracting much attention recently since many underwater vision tasks are relying on the acquisition of clear underwater images. Inspired by the success of deep convolutional neural networks (CNNs) for many high-level vision tasks, in this paper, we propose a multi-scale feature fusion based neural network for underwater image enhancement. First, multi-scale features, including the local features and global features, are extracted. Then, we propose to fuse the global feature with local feature at each scale dynamically. Considering the global features encodes the high-level semantic information and local features holds the structure details at different scales, this fusion strategy is beneficial to underwater image enhancement. Extensive experiments are conducted and the comparisons are made among the state of the art methods as well, where we show great improvements.