A benchmarking toolkit comparing various CPU, OpenMP, and CUDA implementations of classical image denoising algorithms, specifically targetting Monte-Carlo render noise.
To build the project, run the CMake configuration and build commands from the root directory:
cmake -B build -S . -DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc
cmake --build buildThis will configure CMake and compile the executable into the build directory.
Run the GUI or command line utility using the compiled binary.
To launch the interactive split screen interface, run the executable with no arguments (or explicitly pass --gui):
./build/bigdnTo run in CLI mode, specify a filter using the -f or --filter flag:
./build/bigdn -f <filter_name> [options]-h, --help #Show help message and exit
-f, --filter <name> #Filter to run: mean, gaussian, guided, joint_guided, atrous, masked_atrous, or benchmark (default: benchmark)
-d, --device <name> #Execution device/implementation: cpu, omp, or cuda (default: cuda)
-i, --input <path> #Noisy beauty input image path (default: TEST_IMAGES/NOISY.png)
-n, --normal <path> #Normal map input image path (default: TEST_IMAGES/NORMAL.png)
-a, --albedo <path> #Albedo map input image path (default: TEST_IMAGES/ALBEDO.png)
-o, --output <path> #Denoised output image path (default: denoised.png)
-g, --gt <path> #Clean ground truth image path for metric computation (default: TEST_IMAGES/GROUNDTRUTH.png)
-k, --kernel-size <N> #Kernel size (must be a positive odd integer, default: 15)
-p, --passes <N> #Number of A Trous passes (default: 5)
-sc, --sigma-color <V> #Color sigma parameter (default: 0.15)
-sn, --sigma-normal <V> #Normal sigma parameter (default: 0.1)
-sa, --sigma-albedo <V> #Albedo sigma parameter (default: 0.05)
-mk, --median-kernel <N> #Median kernel size (3 or 5, default: 3)
-mt, --median-threshold <V> #Median outlier threshold (default: 0.05)
-r, --runs <N> #Number of iterations to run during benchmark mode (default: 5)Ex- To run the CUDA A Trous denoiser on custom images:
./build/bigdn -f atrous -i TEST_IMAGES/NOISY.png -n TEST_IMAGES/NORMAL.png -a TEST_IMAGES/ALBEDO.png -o output.png -p 5 -sc 0.15 -sn 0.1| Filter / Implementation | Median Time | Speedup vs CPU (ST) |
|---|---|---|
| Mean (Single-Threaded) | ~3488.67 ms | 1.00x |
| Mean (OpenMP) | ~98.99 ms | 35.24x |
| Mean (CUDA) | ~1.11 ms | 3144.42x |
| Gaussian (Single-Threaded) | ~1133.40 ms | 1.00x |
| Gaussian (OpenMP) | ~106.92 ms | 10.60x |
| Gaussian (CUDA) | ~1.13 ms | 1003.00x |
| Guided (Single-Threaded) | ~230.45 ms | 1.00x |
| Guided (OpenMP) | ~125.75 ms | 1.83x |
| Guided (CUDA) | ~5.20 ms | 44.30x |
| Joint Guided (CUDA) | ~6.08 ms | — |
| À-Trous Wavelet (CUDA, 5 passes) | ~12.36 ms | — |
| Masked Median + À-Trous (CUDA) | ~44.20 ms | — |
Quality metrics obtained relative to the clean reference image (TEST_IMAGES/GROUNDTRUTH.png). Higher values for both PSNR and MS-SSIM indicate better denoising quality:
| Implementation | PSNR (dB) | MS-SSIM |
|---|---|---|
| Mean Filter | 28.45 dB | 0.9937 |
| Gaussian Filter | 33.52 dB | 0.9985 |
| Guided Filter | 34.23 dB | 0.9978 |
| Joint Guided Filter (CUDA) | 34.92 dB | 0.9990 |
| À-Trous Wavelet Filter (CUDA) | 30.24 dB | 0.9963 |
| Masked Median + À-Trous Filter (CUDA) | 33.08 dB | 0.9989 |
- Classical edge preserving filters (Guided, Joint Guided, A Trous) have a hard mathematical limit distinguishing high frequency noise from true details, which can leave splotchiness in high noise renders.
- The Joint Guided and A Trous filters are only supported on the CUDA device.
- Standard A Trous does not denoise salt and pepper noise well because pixel outlier spikes are treated as edges and are not smoothed. Resolved by using the Masked Median + A-Trous Wavelet Filter, which pre-filters outlier spike pixels with a selective median filter before using variance-guided A Trous denoising.
- Joint Guided filter is prone to smudginess and detail loss in low-feature areas.
- Gaussian filter is too smudgy and completely blurs geometric edges.