Two-Stage Network Architecture for Sonar Image Denoising Based on U-Net

With the increase of marine activities, underwater target detection has become a key technology for the development of marine resources and maintenance of rights and interests. Due to the complexity of underwater environment, the multiplicative scattering noise in sonar images often affects the accuracy of detection algorithms. Meanwhile, in order to solve the problem of no clean reference image pairing, a two-stage U-Net denoising algorithm (TUDA) is proposed in this paper. First, the method in this paper uses a two-stage architecture to gradually learn the sonar denoising features in the ideal state. Then, the residual encoder-decoder architecture learns the contextual features, and the information exchange between different stages is enhanced by passing the features from the previous stage to the next stage for aggregation through the cross-stage feature fusion and supervised attention module. The experimental results on sonar datasets show that TUDA achieves better results on the no-reference image quality evaluation metrics NIQE, BRISQUE when compared with the traditional and state-of-the-art methods. Therefore, TUDA can better retain the image's texture information and edge structure features while denoising, showing better visual effects.