MSFFNet: Multi-stream feature fusion network for underwater image enhancement

Abstract

Deep learning-based image processing methods have achieved remarkable success in improving the quality of underwater images. These methods usually extract features from different receptive fields through downsampling operations, and then enhance underwater images through upsampling operations. However, these operations of downsampling and upsampling inevitably disrupt the relations of neighboring pixels in raw underwater images, leading to the loss of image details. Given this, a multi-stream feature fusion network, dubbed MSFFNet, is proposed to enrich details, correct colors, and enhance contrast of degraded underwater images. In MSFFNet, the multi-stream feature estimation block is carefully constructed, which separately takes original resolution feature maps and low-resolution feature maps as inputs. The multi-stream feature estimation block proficiently preserves the details information of the original underwater image while extracting high-level features. Besides, a coordinate residual block is designed to emphasize valuable features and suppress noises based on position knowledge. A local–global feature fusion block is presented for selectively fusing the complementary multi-scale features. Finally, extensive comparative experiments on real underwater images and synthetic underwater images demonstrate that the proposed MSFFNet has superior performance on underwater image enhancement tasks.