search

JWSoh / MZSR

Meta-Transfer Learning for Zero-Shot Super-Resolution (CVPR, 2020)
270 stars 62 forks source link

How can you get the kernel? #8

Open lz666win opened 4 years ago

lz666win commented 4 years ago

Nice job for solve the time problem of ZSSR! I notice you have a kernel in input, for SR problem, kernel is important, but you do not provide any code to get kernel. How can you get the kernel?

JWSoh commented 4 years ago

For the synthetic dataset, we have used the ground truth kernel. For real dataset, we have used a kernel estimation algorithm based on MAP framework. Also, you may refer to the very recent one from NeurIPS 2019 "Blind Super-Resolution Kernel Estimation using an Internal-GAN."

zerodohero commented 4 years ago

For the synthetic dataset, we have used the ground truth kernel. For real dataset, we have used a kernel estimation algorithm based on MAP framework. Also, you may refer to the very recent one from NeurIPS 2019 "Blind Super-Resolution Kernel Estimation using an Internal-GAN."

Your work is very admirable I estimate blur kernel from the real image, can I directly replace it with kernel.mat and then use it for super-resolution reconstruction of my real image?There are pre-training models in the program, and they correspond to the parameters of the pre-training model generated by different blur kernel respectively. If I use ablur kernel that does not exist in the pre-training, will the results become worse? Can I use the blur kernel I estimated to retrain the model to get a model corresponding to this blur kernel? Will this improve the results theoretically?