Face Super-Resolution via Meta-Learning Optimization Network

Most current face super-resolution (FSR) methods assume that the degradation from high-resolution (HR) image to low-resolution (LR) image is single and known (e.g., bicubic downsampling), which results in the generalization ability of the model being limited to the ideal conditions for bicubic downsamling. When the actual degradation is different from the current settings, these methods will suffer serious performance degradation. In order to solve this problem, we proposed a face super-resolution via meta-learning optimization network (MLON). Firstly, learning the SR tasks composed of multiple blur kernels to construct an initial SR model with sensitive and transferable parameters; then, the model parameters are optimized with a few gradient steps by learning a new blur kernel; finally, the generalization ability of the current model to the new blur kernel is verified. A large number of experimental results on FEI datasets show that our MLON method is superior to the recent advanced FSR methods.