[Pre-print] [Project Page (Coming soon)]
This repository contains the source code used in experimental evaluations for the paper "NIF25: Reducing the Gap in INR-Based Image Compression".
The Python environment can be handled using conda:
conda env create -f environment.yaml
The command above will create an environment named nif25, activated by:
conda activate nif25
To encode a single image run the encode.py scripts:
python3 encode.py <CONFIGURATION_PATH> <INPUT_IMAGE> <COMPRESSED_IMAGE>
For example, to encode the image #3 from the Kodak dataset located at "test_images/kodak/3.png" and save it at "kodim03.nif" run:
python3 encode.py configurations/nif/kodak/120.yaml test_images/kodak/3.png kodim03.nif
Analogously, you can decode a compressed image using the decode.py script:
python3 decode.py <CONFIGURATION_PATH> <COMPRESSED_IMAGE> <DECOMPRESSED_IMAGE>
To decompress the image encoded in the example above, use:
python3 decode.py configurations/nif/kodak/120.yaml kodim03.nif kodim03_decoded.png
Input files should be encoded in lossless PNG format. Output file names end with ".nif" by convention, but this is not forced by code.
To reproduce the experiments discussed in the paper, run the corresponding script in the folder "scripts/datasets/". These scripts will generate a bash file named "schedule.sh", filled with the sequence of commands that will encode a whole dataset. Logs will be saved under "logs/" and compressed files, along decompressed images and statistics will be saved under "results/".
For instance, to generate a "schedule.sh" to reproduce experiments on Kodak use:
./scripts/datasets/kodak.sh
Then execute it with ./schedule.sh.
If you want to perform a full encode-decode experiment on a single image and to export stats, on Linux/MacOS systems you can use the "scripts/run/full_experiment.sh" script as follows:
./scripts/run/experiment.sh <CONFIGURATION_PATH> <INPUT_IMAGE> <OUTPUT_FOLDER>
The compressed and then decoded images, along with stats saved in JSON, will be exported in the indicated output folder.
The kodak dataset is included in this repository, along with compressed ".nif" files that can be decoded using this software.
The uncompressed CLIC2020 validation images used in the experiments can be download at this link.