A Learning-based Projected Gradient Descent Approach for Blind Super-resolution
Zhixiong Yang, Huaizhang Liao, Han Zhang, Jingyuan Xia · 2022
Kernel estimation is an important component in the blind super resolution (BSR), and significantly dominates the performance of the restored image. Most of the state-of-the-art approaches, such as FKP-DIP, is proposed to formulate the kernel estimation as a cumbersome network-based end-to-end model, which provides promising performance through kernel prior pre-training behavior but lack of flexibility and generalization-ability towards different kernels in terms of scales, categories and blur level. Thus, this paper strives to reformulate the kernel estimation through a learning aided projected gradient descent (LPGD) algorithm, referring to LPGD-DIP. The LPGD algorithm iteratively optimizes the kernel with respect to the gradient based principle, therefore providing good generalization-ability for arbitrary kernel estimation. Meanwhile, a lightweight network is applied to dynamically and adaptively learn the update rule at each step, whereby the performance is also guaranteed through the learning-based network behavior. Numerical results show that the LPGD-DIP surpass the existing BSR approaches in the aspect of the kernel estimation, the image reconstruction in the most Gaussian blur scenarios with a modest computational complexity. Our code is available at https://github.com/XiaGroup/LPGD-DIP.