Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission

Abstract

Due to the complexity of underwater imaging environments, images captured via optical vision systems often exhibit significant degradation. To combat this issue, we introduce a multi-scale feature fusion underwater image enhancement network, termed MFUNet. MFUNet is a novel multi-scale feature fusion network, guided by medium transmission, ensures the content integrity of the reconstructed image by leveraging interaction features among non-adjacent layers. This approach addresses the common problem of the loss of image detail features. Moreover, MFUNet enhances the response to high-frequency information by employing edge loss, thereby improving sensitivity to edges and textures. By deepening the network hierarchy, the image undergoes deep encoding and decoding, which maximizes the multi-color space encoder's and multi-scale feature fusion's potential in color representation and enhances the structural similarity and overall quality of the image. It is worth noting that we achieved superior performance by utilizing fewer model parameters. Extensive experiments across various datasets demonstrate that our method surpasses comparative methods in both visual quality and quantitative metrics.