Unsupervised image denoising based on self-attention mechanism
Qingmeng Guo, Weibo Wei, Nianrun Xi, Bo Wang · 2023
Unsupervised image denoising has emerged as a critical research field for reducing noise in images in recent years. Unsupervised models solve the problem of high costs associated with collecting training samples and demonstrate greater adaptability to noise. This paper proposes an improved unsupervised image denoising model called IZS-N2N (Improved Zero-Shot Noise2Noise) based on ZS-N2N (Zero-Shot Noise2Noise). The network structure of the ZS-N2N model is overly simplistic and the loss function is suboptimal. To address these issues, the model incorporates several improvements. Firstly, it combines image downsampling and convolutional neural network in ZS-N2N to carry out denoising iteration. Additionally, the network layer is redesigned to include global context modeling and a channel attention mechanism. Furthermore, the IZS-N2N model modifies the loss function to accommodate noises of different magnitudes. The network consists of five layers and contains only 23,000 parameters, which enhances its lightweight nature. The experiment was conducted using two parts: visual subjective evaluation and PSNR objective evaluation. The experimental results demonstrate that this model improves the denoising details compared to current mainstream unsupervised denoising models. Additionally, the objective indicators' evaluation values are also at a satisfactory level.