BHI_LR_multi / README.md
Phips's picture
Update README.md
005b77e verified
|
raw
history blame
1.25 kB
metadata
license: cc-by-4.0

BHI_LR_multi

This is a x4 LR counterpart to the BHI SISR Dataset.

SCREENSHOT HERE

To create this dataset, a 0.25 scaling (or x4 scaling) has been applied, with multiple scaling algorithms in a randomized manner.

The algos used are linear, cubic_mitchell, lanczos, gauss, box and down_up (with these algos). The specifically applied degradation for each image can be found in the applied_degradations.txt file

The trained model learns to handle multiple scalings, and also these counteract each other in the sense of that lanczos will always sharpen, gauss will always soften etc.
Using this instead of just bicubic sampling for example will help the model not to be trained on / pick up on one specific scaling algos characteristics.

screenshot beginning of applied_degradations.txt
Applied scaling on each image, screenshot beginning of applied_degradations.txt