This repository includes code for the PatchSVD algorithm introduced in PatchSVD: A Non-uniform SVD-based Image Compression Algorithm. PatchSVD is a simple image compression algorithm that improves SVD-based image compression by applying non-uniform compression to different patches in the input image utilizing 1-rank SVD heuristics. When we need to compress images that contain sharp changes in pixel intensity, such as compressing images that contain text, PatchSVD outperforms JPEG.
After creating your virtual environment, install the dependencies by running the following command:
pip install -r requirements.txt
Then, you can compress any image:
cd patchsvd
python compression_experiments.py --img-path $IMAGE_PATH --p_x $p_x --p_y $p_y --target-compression $CR --output-dir $output_dir
If you wish to save the PatchSVD visualizations that shows the complex and simple patches, pass the --visualize flag to the above command. You can set a visualization limit to only save the first few visualizations by setting --visualization_limit $limit.
You can run PatchSVD on a dataset by passing --dataset $dataset_name instead of --img-path as follows:
cd patchsvd
python compression_experiments.py --dataset $dataset_name --p_x $p_x --p_y $p_y --target-compression $CR --output-dir $output_dir
Currently, the supported datasets are MNIST, CIFAR-10, FGVC_Aircraft, EuroSAT, Kodak, and CLIC; but you can add support for any dataset you wish following the examples in compression_experiments.
When you run compression_experiments.py, a csv file will be generated as well as a folder that contains image outputs. The csv file contains the calculated metrics that can be used later for plotting.
This repository includes code that visualizes how PatchSVD patches are categorized, as well as code that creates the plots that compares PatchSVD with SVD and JPEG. All the comparison plots are illustrated in the paper.
Please cite the following paper if you used PatchSVD in your work:
Golpayegani, Z., & Bouguila, N. (2024).
"PatchSVD: A Non-uniform SVD-based Image Compression Algorithm."
In Proceedings of the International Conference on Pattern Recognition Applications and Methods (ICPRAM).