AR-CNN, Fast AR-CNN

Credits: https://github.com/yjn870/ARCNN-pytorch

This repository is implementation of the "Deep Convolution Networks for Compression Artifacts Reduction".
In contrast with original paper, It use RGB channels instead of luminance channel in YCbCr space and smaller(16) batch size.

Requirements

• PyTorch
• tqdm
• Numpy
• Pillow

Usages

Train

Data augmentation option --use_augmentation performs random rescale and rotation.

python main.py --arch "ARCNN" \     # ARCNN, FastARCNN
               --raw_images_dir "" \
               --comp_images_dir "" \
               --val_raw_images_dir "" \
               --val_comp_images_dir "" \
               --outputs_dir "" \
               --patch_size 24 \
               --batch_size 16 \
               --num_epochs 20 \
               --lr 5e-4 \
               --threads 8 \
               --seed 123      

Test

Output results consist of image compressed with JPEG and image with artifacts reduced.

python inference.py --arch "ARCNN" \     # ARCNN, FastARCNN
               --weights_path "" \
               --image_path "" \
               --outputs_dir "" \