Mini-OpenCV — GPU Image Processing Library

English | 简体中文

A high-performance CUDA-based image processing library providing GPU-accelerated operators for computer vision applications.

⚡ Performance: 30-50x faster than CPU OpenCV for comparable operations

Tested on RTX 4090 with 4K images vs OpenCV 4.8 CPU implementation. See benchmarks/ for details.

---

📚 Quick Links

Resource	Description
Installation	Complete setup guide
Quick Start	Get started in 5 minutes
API Reference	Complete API documentation
Examples	Code examples and tutorials
Specifications	Requirements, RFCs, and technical designs

Full Documentation: https://lessup.github.io/mini-opencv/

---

✨ Features

Category	Operators	Highlights
Pixel Operations	Invert, grayscale, brightness	Per-pixel parallel
Convolution	Gaussian blur, Sobel edge detection	Shared memory tiling
Histogram	Calculation, equalization	Atomic ops + reduction
Geometric	Rotate, resize, flip, affine	Bilinear interpolation
Morphology	Erosion, dilation, open/close	Custom structuring elements
Threshold	Global, adaptive, Otsu	Histogram-driven
Color Space	RGB/HSV/YUV conversion	Batch conversion
Filters	Median, bilateral, sharpen	Edge-preserving
Pipeline	Multi-step async processing	Multi-stream concurrency

---

🏗️ Architecture

See Architecture Specification for complete design details.

Three-layer design: High-level APIs → CUDA Kernels → Infrastructure

Application Layer → Operator Layer → Infrastructure Layer

---

🚀 Quick Start

# Clone and build
git clone https://github.com/LessUp/mini-opencv.git
cd mini-opencv
cmake -S . -B build -DBUILD_EXAMPLES=ON
cmake --build build -j$(nproc)

# Run tests
ctest --test-dir build --output-on-failure

# Run example
./build/bin/basic_example

Basic Usage

#include "gpu_image/gpu_image_processing.hpp"
using namespace gpu_image;

// Step 1: Create an ImageProcessor
ImageProcessor processor;

// Step 2: Create/Load a host image
HostImage hostImage = ImageUtils::createHostImage(1920, 1080, 3);
// Fill hostImage.data with your pixel data (RGB, 8-bit per channel)

// Step 3: Upload to GPU and process
GpuImage gpu = processor.loadFromHost(hostImage);
GpuImage blurred = processor.gaussianBlur(gpu, 5, 1.5f);
GpuImage edges = processor.sobelEdgeDetection(gpu);

// Step 4: Download result back to host
HostImage result = processor.downloadImage(edges);

---

📋 Requirements

System Requirements

Component	Minimum	Recommended
CUDA	11.0	12.x
CMake	3.18	3.24+
C++	C++17	C++17
GPU	CC 7.5+ (Turing)	RTX 30/40 series

Supported Image Formats

Format	Read	Write
JPEG/JPG	✓	✓
PNG	✓	✓
BMP	✓	✓
TGA	✓	✗

Note: All formats support 8-bit per channel (Grayscale, RGB, RGBA)

---

📖 Documentation

Complete documentation available at GitHub Pages:

• Installation Guide
• Quick Start
• Architecture Overview
• Performance Optimization
• API Reference
• FAQ

Specifications

Technical requirements and design documents:

• Product Requirements
• Architecture Design
• API Reference
• Architecture Overview

---

🤝 Contributing

See CONTRIBUTING.md for guidelines.

---

📝 Changelog

See CHANGELOG.md for version history.

---

🛠️ Troubleshooting

Q: CMake cannot find CUDA / nvcc not found

# Set CUDA path explicitly
export CUDAToolkit_ROOT=/usr/local/cuda
# Or when configuring:
cmake -S . -B build -DCUDAToolkit_ROOT=/usr/local/cuda

Q: Running examples shows "CUDA is not available"

• Ensure you have an NVIDIA GPU with Compute Capability 7.5+
• Install CUDA Toolkit 11.0+ from NVIDIA
• Verify: nvidia-smi should show your GPU

Q: Tests fail or crash during execution

• Check GPU memory: Large images may require more VRAM
• Try smaller test images or reduce batch size
• Run with ctest --test-dir build --output-on-failure -V for verbose output

Q: How to verify installation is successful?

cd build && ctest --output-on-failure
# All tests should pass

For more issues, check GitHub Discussions.

---

📄 License

MIT License — see LICENSE file.

---

⭐ Star this repo if you find it helpful!

For support, open an issue or start a discussion.