Dejpeg GAN

Chainer implementation of a GAN for recovery of data lost during JPEG compression. Also includes batch testing and statistics collection scripts, and a script to visualize this data in graphs and numbers. Adapted from superresolution model: https://github.com/Hi-king/superresolution_gan

Training Without GPU

python train.py --train_data_paths "/path/to/images/*.jpg" --model_output_path "/path/to/save/trained/models"

Note: the output directory must not exist, it will be created during runtime.

Training With GPU

python train.py --train_data_paths "/path/to/images/*.jpg" --model_output_path "/path/to/save/trained/models" --use_gpu=0

Note: cupy must be installed and set up. You will need to identify your GPU device ID. In this example, it is 0. You may simply try numbers from 0 upwards until one works.

Testing Individual Images with Individual Models

python test.py --model_path "/path/to/model.npz" --input_path "/path/to/input/image.jpg" --output_path "/path/to/output/image.png" --use_gpu=0

Note: you may omit the output_path argument and the results will be displayed in a window instead.

Testing Individual Images with Multiple Models

python test_single.py --models_directory_path "/path/to/models/directory" --input_file_path "/path/to/input/image.jpg" --output_directory_path "/path/to/output/generated/images" --use_gpu=0

Testing Batches of Test Images with Multiple Models

python test_batch.py --models_directory_path "/path/to/models/directory" --input_directory_path "/path/to/test/images" --output_directory_path "/path/to/output/generated/images" --use_gpu=0

Note: consists of 3 steps: preprocessing original images to input images, then generating all model output images, then computing statistics across all images. Additional flags are available to skip any step(s) in the process: --do_preprocessing=true, --do_output_generation=true, --do_performance_evaluation=true

python display_results.py --statistics_file_path "/path/to/statistics/file.pkl"

Note: the .pkl file will be generated after the final step of the previous script above this one. Graphs displayed can be customized in this script itself.

Image Manipulation

To convert images between formats:

convert input.jpg output.png

To change image resolution:

mogrify -resize 50% image.jpg

To compress to specified JPEG compression level:

convert input.png -quality 10 output.jpg