DF2K_BHI SISR Dataset

This is a filtered version of the DF2K dataset, specifically aimed for single image super-resolution model training.
It is a result, or is one of the testing sets, for my huggingface community blog post on my BHI-Filtering method.

Dataset Details

This is based on the DF2K dataset, which consists of 3450 images.

These images then have been tiled into 21'387 512x512px tiles for faster sisr I/O training speed.

The BHI filtering has been applied with the following default threshold values:

Blockiness < 30
HyperIQA >= 0.2
IC9600 >= 0.4

Which results in this DF2K_BHI dataset with 12'639 512x512 training tiles.

Visualization

Here is a visual example of the first few tiles that are in the dataset:

Visual example - first few tiles of the DF2K_BHI Dataset

Training

As can be seen in the following training validation tensorboard graphs (DIV2K validation dataset), this BHI-filtered version of the DF2K dataset is able to achieve higher or similiar metric scores while seeing a reduction of -40.9% in training tiles quantity.

Here is the training validation metrics on the DIV2K validation dataset, using the plksr_tiny architecture option, in comparison to the tiled full DF2K dataset:

Tiled DF2K dataset (here called X): 21'387 tiles
DF2K_BHI dataset: 12'639 tiles

Tensorboard: SSIM validation scores of DF2K and DF2K_BHI

Tensorboard: PSNR validation scores of DF2K and DF2K_BHI

Tensorboard: DISTS validation scores of DF2K and DF2K_BHI