LOCAL TO GLOBAL PRIOR LEARNING FOR BLIND UNSUPERVISED IMAGE SUPER RESOLUTION

Deep convolutional neural networks (DCNN) have dominated the single image super resolution (SR) field, and demonstrated significant success in generating high-resolution (HR) images from the ideal low-resolution (LR) images captured under controlled imaging conditions. However, the resolved image performance would be dramatically degraded for the LR image captured in real environment. Recently, some works attempt to automatically learn image and kernel priors by leveraging the local statistic modeling capability of the DCNNs from the LR observation only, and illustrated the feasibility to construct a specific CNN for real-world application. The basic convolution operation in the DCNNs can potentially capture the local interaction to generate plausible image appearance but are far from sufficiency to achieve global context, which has been proven to be a promising property of a transformer block, to further boost the SR performance. This study proposes a cooperative local to global prior learning (LoGPT) framework for blind unsupervised image super resolution by jointly modeling the local connectivity with convolution operations and global context with transformer block. Specifically, we elaborate a multi-scale encoder-decoder architecture configuring with the convolution blocks on the high-resolution scales to learn local priors while the transformer block on the low-resolution scales to capture long-range dependencies, and then incorporate with a simple convolution-based subnet to simultaneously learning the local to global image priors and kernel priors in an unsupervised way using the observed LR image only. Extensive experiments have demonstrated that our proposed blind SR method achieves superior SR performance over both supervised and unsupervised state-of-the-art methods in term of quantitative metrics and perceptive quality.