🖼️ Optimize Image Segmentation with Parallel KMeans Clustering

• Developed 4 versions of hybrid MPI and CUDA-based KMeans clustering to accelerate image segmentation on 100 pet images
• Converted each image into up to 2,073,600 RGB pixels using OpenCV and distributed across 2 GPUs in C++, parallelizing distance calculations and centroid updates

Members

• Minh Le
• Truong Dang
• Dr. Andrew Sohn

💭 Inspiration

Why Image Segmentation?

Help isolate subjects (like pets in our test images) by removing background clutter and unnecessary features. This process also helps improve tasks such as pet recognition (object detection), classification, or focused editing.

Why KMeans?

A simple data-driven algorithm that works well with multi-dimensional features (3.g. RGB is in 3-dimension).

Why Parallelize?

As real-world images (Full HD, 4K, Digital Camera) contain millions of pixels, we implement KMeans in plain C++, MPI, and CUDA to compare their performance and learn HPC patterns.

This project is the intersection of computer vision, ML, and HPC, providing an understanding of scalable segmentation.

🛠️ How To Run

1. Create a build directory and run CMake

mkdir build
cd build
cmake ..
cmake --build .

2. Run the Executable

Executables are located in the build/bin folder:

• Serial version: ./bin/img_seg_serial
• MPI version: ./bin/img_seg_mpi
• CUDA version: ./bin/img_seg_gpu
• MPI + CUDA version: ./bin/img_seg_mpi_cuda

ℹ️ Navigate to the build folder to run the executables.

Serial & CUDA versions:

<executable> <input-path> <num-clusters> <output-path>

MPI & MPI + CUDA versions:

mpirun -np <numprocs> <executable> <input-path> <num-clusters> <output-path>

3. Commands Used to Test

Serial

./bin/img_seg_serial ../db/input/dog48.jpg 5 ../db/output/output_dog48_serial5.jpg

./bin/img_seg_serial ../db/input/cat2.jpg 5 ../db/output/output_cat2_serial5.jpg

./bin/img_seg_serial ../db/input/dog35.jpg 7 ../db/output/output_dog35_serial7.jpg

MPI

mpirun -np 4 ./bin/img_seg_mpi ../db/input/dog48.jpg 5 ../db/output/output_dog48_mpi5.jpg

mpirun -np 4 ./bin/img_seg_mpi ../db/input/cat2.jpg 5 ../db/output/output_cat2_mpi5.jpg

mpirun -np 4 ./bin/img_seg_mpi ../db/input/dog35.jpg 7 ../db/output/output_dog35_mpi7.jpg

CUDA

./bin/img_seg_gpu ../db/input/dog48.jpg 5 ../db/output/output_dog48_cuda5.jpg

./bin/img_seg_gpu ../db/input/cat2.jpg 5 ../db/output/output_cat2_cuda5.jpg

./bin/img_seg_gpu ../db/input/dog35.jpg 7 ../db/output/output_dog35_cuda7.jpg

MPI + CUDA

mpirun -np 4 ./bin/img_seg_mpi_cuda ../db/input/dog48.jpg 5 ../db/output/output_dog48_hybrid5.jpg

mpirun -np 4 ./bin/img_seg_mpi_cuda ../db/input/cat2.jpg 5 ../db/output/output_cat2_hybrid5.jpg

mpirun -np 4 ./bin/img_seg_mpi_cuda ../db/input/dog35.jpg 7 ../db/output/output_dog35_hybrid7.jpg

💻 Technologies

• C++
• MPI
• CUDA
• OpenCV
• GPUs