An implementation of the Edge-Preserving Image Smoothing (EPIS) algorithm discussed in this paper.
Our goal is to measure the impact of L1 piece-wise flattening with edge-preserving on image compression for various image formats.
EPIS transforms an image (left) into a flattened image (right).
python epis.py [filepath to image]- Speedup
compute_pairs()by generating the matricespair1andpair2instead building them iteratively.
Libraries used and their respective versions. The minimum versions for each library are unknown. A detailed list can be found in library_versions.txt.
- Python 3.10.0
- Numpy 1.24.2
- CuPy 12.0.0
- CuSPARSE 0.4.0
- CUDA 12.1
The hardware used for test_data are:
- AMD Ryzen 5 5600x 6-Core 12-Thread
- 8GB Nvidia RTX 3070
- 32 GB DDR4-3600 RAM
EPIS relies heavily on GPU computation, specifically sparse matrix multiplication and sparse matrix solving. EPIS uses incredibly large sparse matrixes which necessitates a large pool of VRAM. This is the limiting factor when flattening images. Downscaling images is required depending on the hardware used.
A GPU with 8GB can flatten images with dimensions of about 180x180 pixels. An example hardware usage analysis on architecture_118_180.png is shown below.
