Unsupervised learning for blind single image super-resolution

We will provide implementation of this method with TensorFlow version code.

Some experiment results on real-world LR image SR:

As shown in Figure 1, our model can generate high-resolution image with less artifacts and more details.

Figure 1

We have evaluated our approach on 400 real-world LR images. Figure 2 shows the comparison results for six (more detailed comparison can be found in the supplementary material): chip (244 × 200), cat (429 × 380), comic (358 × 578), flower (334 × 490), store (440 × 430) and building (440 × 430). Image cat, comic, flower, store and building are stored in JPEG format, unknown quantization-related distortions (e.g., block artifacts) are inevitable.

Figure 2

TODO list

• update Readme
• upload my code
  • add super parameters
• upload more results