###### tags: `Paper Study`

# Super Resolution

### [2020] Closed-loop Matters- Dual Regression Networks for Single Image Super-Resolution

**Highlights**

1. **Existing SR Problems**: (1) There exist infinite HR images that can be downsampled to the same LR image. Thus the space of the possible functions that map LR to HR images becomes extremely large. (2) SR methods rely on paired data. LR-HR data may be unavailable in real-world.The underlying degradation is unknown. (The real-world data do not have the same distribution to LR images obtained from degradation method, e.g. Bicubic)

2. Dual regression training: (1) Add constraint to LR images (SR images reconstruct LR images) (2) No need HR data. Directly train from the LR images in real-world.


3. Dual regression training for paired data:


4. Dual regression training for unpaired data:


5. Architecture


6. TODO results, model size

**Ideas**

1. Dual regression for MFSR/SISR.


2. Dual regression + Crop segmentation ROI for SISR.