Underwater Image Enhancement Based on Multi-Scale Feature Fusion and Attention Network

Abstract: Underwater image plays an important role in the exploration of marine resources. Aiming at the problems of incomplete defogging and loss of details in existing underwater enhancement methods, we propose a method based on multi-scale feature fusion attention network. Firstly, the multi feature extraction module extracts image features, learning the different space feature information. Secondly, we use the feature fusion module to strengthen the connection of different spatial information and realize the reuse of features. Thirdly, a feature modulation module is constructed to transform low-quality information features into high-quality ones, including channel and pixel attention residual blocks, which are stacked into a chain structure. Multi-level features are dynamically modulated to enhance image details and suppress redundant information. Finally, constructing a polynomial loss function contains a mean square error loss function, loss function and perceptual loss function. Additionally, the asynchronous training mode is introduced to improve network performance. The comparative experiment shows that based on the EUVP dataset, synthetic SUDS dataset and UFO-120 dataset, the proposed method is superior to other classical and novel methods in subjective visual quality and objective evaluation indicators (UCIQE, NIQE, SURF and information entropy). The enhanced underwater image has an excellent defogging effect and presents conspicuous advantages in restoring image details, which significantly improves the visual quality of the underwater image.