Adaptive Non-Local Regression Prior based on Transformer for Image Deblurring
Yixing Ji · 2023
The non-local regression prior has shown promising results in image deblurring by effectively combining local continuity and non-local self-similarity of images. However, traditional non-local regression priors have limited representation capacity, and existing non-local regression-based methods often require iterative solutions to complex optimization problem. Additionally, these methods involve manually tuning hyper-parameters, which can be difficult and time-consuming. To address these limitations, we propose an adaptive non-local regression prior based on the Transformer model for image deblurring. Different from other learning-based approaches, the self-attention mechanism in the Transformer module better aligns with the modeling process of non-local regression prior, and the global computation capability of Transformer enables effective modeling of global context information. By unfolding the iterative process of the model into a neural network, we introduce a novel deep unfolding network. The proposed network is trained end-to-end, allowing all parameters to be jointly optimized to facilitate image restoration. The extensive experiments show that our proposed model outperforms state-of-the-art model-based and learning-based methods both in terms of PSNR metrics and visual quality.