CUDA Accelerated Box Filter for Texture Image Enhancement

This project applies a CUDA-accelerated box filter to a large set of monochrome texture images from the USC-SIPI Image Database - Textures Volume. The objective is to demonstrate high-throughput GPU processing of real-world image datasets, with a focus on scalability, performance, and batch automation.

🚀 Project Overview

• Purpose: Enhance texture images by applying a uniform box filter to reduce noise and improve clarity.
• Dataset: 64 TIFF grayscale images from the USC-SIPI texture collection.
• Technology: NVIDIA CUDA, C++14, OpenCV
• Kernel: Custom box filter implemented using shared memory for efficiency.

🧪 Requirements

• CUDA Toolkit (tested on CUDA 11+)
• OpenCV 4.x (libopencv-dev)
• C++14 compiler
• Linux system with NVIDIA GPU (sm_61+ recommended)

Install OpenCV (if not already installed):

sudo apt-get install libopencv-dev

🔧 Build Instructions

Compile using the provided Makefile:

make

Or manually:

nvcc -O2 -std=c++14 `pkg-config --cflags opencv4` \
    -Iincludes src/main.cu src/image_utils.cpp \
    -o textures_cuda `pkg-config --libs opencv4` -lcudart

▶️ Usage

Run the binary and specify the image folder and output path:

./textures_cuda <input_folder> <output_folder> <filter_width>

Example:

./textures_cuda textures outputs 5

• input_folder: Path to folder with .tiff texture images
• output_folder: Where to save processed images
• filter_width: Size of the box filter kernel (must be odd, e.g., 3, 5, 7)

You can also use the helper script:

bash run.sh

📷 Dataset Download

The images are from the USC-SIPI Textures Volume. Download the zip or tar:

wget http://sipi.usc.edu/database/zip/textures.zip
unzip textures.zip -d textures/

✅ Proof of Execution

A sample set of before-and-after .tiff images, processing logs, and batch summaries are included.

Examples:

Original Image	Filtered Image
images/1.1.01.tiff	output/1.1.01.tiff

---

Made by Aditi Saxena