Non-Local Means (NLM) Denoising

An efficient image denoising implementation featuring Serial CPU, standard CUDA, and optimized CUDA Shared Memory versions.

Results

This project was benchmarked on Perlmutter High Performance Computing (HPC) nodes.

• /results/ contains the raw data generated on the HPC nodes.
• /results/images/ contains the visual outputs and performance plots (e.g., speedup.png, runtime-GPU.png).

---

1. Prerequisites

• Hardware: NVIDIA GPU (for CUDA implementations).
• Software: NVIDIA CUDA Toolkit, GCC, and MATLAB (for image conversion).

2. Configuration

Before running, open each .c or .cu file and adjust the global variables at the top of the script:

• PIXELS: The dimension of your square image (e.g., 512 for 512x512).
• PATCH_SIZE: The size of the comparison window (typically 3-7).
• FILTER_SIGMA: Controls the smoothness of the result.
• PATCH_SIGMA: Controls the sensitivity to patch similarity.

Note: Higher sigma values result in more aggressive denoising but risk blurring fine details.

3. Workflow

The workflow requires MATLAB to convert images to text files, the C/CUDA code to perform the filtering, and MATLAB again to reconstruct the PNG output.

Step-by-Step Execution:

1. Prepare your image: Place your input image (must be square, max 512x512) in the directory and rename it to image.png.
2. Convert to text:

   matlab -batch "run image_read.m"

3. Compile and Run:

   make clean
   make all
   ./nlm-cuda        # Or use ./nlm-serial or ./nlm-cuda-shared

4. Reconstruct output:

   matlab -batch "run denoised_image_read.m"

   This generates filtered_image.png and noise_removed.png.