Meta-learning based blind image super-resolution approach to different degradations

Zhixiong Yang, Jingyuan Xia, Shengxi Li, Wende Liu, Shuaifeng Zhi, Shuanghui Zhang, Li Liu, Yaowen Fu, Deniz Gündüz

Abstract: Highlights•A network-based but pre-training-free blind SISR method, DDSR, is firstly proposed.•DDSR adopts a sampling-based network to replace the pre-training-based networks.•A meta-learning optimization is proposed to prevent trapping into bad local modes.•Extensive simulations validate the superior performance of DDSR compared with SOTAs.