To generate a dataset you need to have RAW 540p videos. Using GenDataset.bash you can convert the videos to frames and generate a dataset.
mkdir Temp/Dir
bash GenDataset.bash Dir/to/videos/ QP Dir/to/destination/ Temp/Dir
Make sure you create 'Temp/Dir/' before running the script. Also, in your videos directory, there should be two folders, train and test, containing .mov files. A the destination dataset will have the same organization.
Train a model using an exisiting dataset by using train.py. The parameters are stated below:
python train.py
--LR_dir [Directory containing RAW, 540p images (QP=0)]
--NR_dir [Directory containing compressed, 540p images (QP!=0)]
--model_save_dir [Directory to save model weights, file name: "model_best_weights.h5"]
Test a model on a set of compressed images by using test.py. Required parameters are shown below:
python test.py
-inr [Directory containing compressed images, QP!=0]
-o [Output directory to save enhanced images]
-m [The model file to use for test (.../model_best_weights.h5)]
To compare the enhanced results (or the not-enhanced compressed frames) with RAW frames and calculate PSNR, PSNR_gen.py can be used.
python PSNR_gen.py
SAVE_FILE [File name to save PSNR results (.mat)]
NOISE_PATH [Directory containing noisy images]
SAVE_PATH [Directory to save PSNR results (.mat)]
SIGNAL_PATH [Directory containing signal (noNoise) images]