Optimizing Adaptive Thresholding Using CUDA for Real-Time Image Processing

📝 Project Overview

This project implements a GPU-accelerated adaptive thresholding technique for real-time image processing using CUDA. The goal is to optimize per-pixel binary thresholding using shared memory and parallelism on NVIDIA GPUs. This method is well-suited for applications in low-latency surveillance and vision systems.

---

⚙️ Dependencies

• CUDA Toolkit: 12.6
• Python (for pre/post-processing scripts): >=3.8
  • numpy
  • Pillow (PIL)

Tested on ASAX GPU servers.

---

📁 Repository Structure

GPGPU_Project/
├── src/
│   ├── checkpoint2/
│     ├── metrices/ 
│     ├── output_images/
│     ├── host_side_implementation.py
│     ├── main.cu
│     └── Makefile 
│   ├── checkpoint3/
│     ├── main.cu 
│     ├── Makefile
│     ├── run_me.sh
│     ├── submission.pbs
│     ├── output_images/
│     ├── optimization/
│       ├── main.cu 
│       ├── Makefile
│       ├── run_me.sh
│       └── submission.pbs
│     ├── metrices/  
├── scripts/
│   ├── jpg_to_raw.py
│   ├── raw_to_jpg.py
├── input_images/
│   └── sample_image.jpg
├── research/
├── requirements.txt
└── README.md 

🛠️ Build Instructions

Navigate to the final optimized checkpoint3:

Baseline implementation

cd src/checkpoint3/

To generate JPG to RAW (Optional, already saved into the input_images directory)

python ../scripts/jpg_to_raw.py image_path --output

Build and run the code using the provided shell file:

./run_me.sh img.raw img.raw.meta

Metrics will be saved in the metrics directory.

Optimized Implementation (APOD)

Build and run the code using the provided shell file:

./run_me.sh img.raw img.raw.meta

Metrics will be saved in the metrics directory.

To generate RAW to JPG

python ../scripts/raw_to_jpg.py raw_image_path meta_path