Generative-Priors-for-Blind-Super-Resolution

Prior information is very important in SR problems, and SR methods with MAP architecture usually require kernel estimation and image priors. Many existing SR methods are based on fixed blur kernel assumptions, such as bicubic downsampling. However, when the actual image does not conform to this fixed assumption, or when the kernel estimation effect is poor in the blind SR problem, the performance of the SR results drops sharply. For the image prior, it is also beneficial to use the network structure itself as a prior method, but it is computationally expensive and impractical in practical applications. To deal with these issues, we propose fully generated prior information and fuse it into the encoder-decoder structure, the Generative Priors for Blind Super-Resolution (GPBSR). Using VAE self-supervision to learn the bijection of the latent space and blur kernel, and the latent space obeys the Gaussian distribution and conforms to the characteristics of the kernel itself, so that the kernel prior information generated by sampling can be well integrated into different SR model frameworks. Similarly, using GANs pre-trained on large datasets can also characterize the prior distribution of the data, with rich and diverse prior information on the input LR map output. This is the research-oriented project of ECE 285 at UC San Diego.