Self-Supervised Image Denoising With a Residual Learning Network
Image denoising through deep learning has seen significant advancements utilizing a large volume of data for network train-ing.However,acquiring clean,noise-free images under real-world conditions remains challenging,leading to the emergence of self-supervised deep learning techniques.Addressing the questions of enhancing self-supervised learning performance and its adaptability across multiple networks,we introduce a self-supervised image denoising approach.Our proposed method involves generating image training pairs by subsampling the noisy image twice and integrating UNet and ResNet to create a robust denoising network.By effec-tively combining these elements,we aim to enhance the denoising capabilities of the latest self-supervised technique,Neighbor2Neigh-bor.Our strategy focuses on constructing a deeper convolutional network while mitigating potential issues like gradient explosion through meticulous parameter tuning.Experimental results demonstrate the efficacy of our approach,showcasing an average improve-ment of 0.3 dB in PSNR value and 0.004 in SSIM value compared to the original algorithm.Additionally,the study confirms the adapt-ability and robustness of the training strategy within the Neighbor 2Neighbor algorithm across various network architectures.
万方数据
维普
