Research on Underwater Image Denoising Based on Dual-Channels Residual Network

YANG Jingjing,XIE Haiyan,XUE Nini,ZHANG Aoming

DOI: https://doi.org/10.19678/j.issn.1000-3428.0064662

2023-01-01

Abstract:As an effective underwater exploration tool,underwater image denoising has received increasing attention and research in academia. To address the vulnerability of traditional filtering methods to image degradation in the denoising process,with obvious noise residues,as well as facilitate the design of filters tailored to specific types of noise,an improved dual-channels residual Convolutional Neural Network（CNN） model is proposed herein on the basis of a deep CNN to eliminate noise from underwater images. The proposed model consists of a Local Residual Network（LRN）,Global Sparse Network（GSN）,and Feature Processing Block（FPB）. The multi-level local noise and global noise features of underwater images are extracted via the dual-channels LRN and GSN in parallel. The channel connection in FPB is then used to fuse the extracted noise features from LRN and GSN,and the convolution layer is employed to enhance the noise information. On this basis,two double loss functions-Mean Square Error（MSE） and Mean Absolute Error（MAE）are used to optimize the network parameters,and the convolution layer is used to reconstruct the underwater images.The experimental results show that compared with that of BM3D,IRCNN,DnCNN,and other methods,the average Peak Signal-to-Noise Ratio（PSNR） of the proposed method increases by 0.02～3.52 dB.Thus,the proposed method successfully removes various levels of random noise and reconstructs clear underwater images.