Mini-OpenCV

v3.0.0 · CUDA 14 · C++17

High-performance CUDA image processing library achieving 30-50× speedup over CPU OpenCV. Supports 9+ operator categories with a three-layer architecture and clean C++17 API. Comprehensive GoogleTest suite and Google Benchmark included.

30-50× Speedup9+ OperatorsMIT License

Core Features

⚡ High Performance

CUDA kernel optimizations: shared memory tiling, atomic operations, warp-level primitives

🧠 Smart Memory

Zero-copy optimization minimizes host-device transfers, memory pool reuse reduces allocation overhead

Zero-Copy Memory Pool Stream Async

🏗️ Three-Layer Architecture

Application → Operator → Infrastructure, clear separation of concerns

ImageProcessor Operators DeviceBuffer

📊 9+ Operators

Convolution, morphology, geometric transforms, histogram, threshold, color space, filters, etc.

Convolution Morphology Geometric

🧪 Well Tested

GoogleTest unit tests + Google Benchmark performance baselines ensure correctness and performance

Benchmarks Examples

📖 Bilingual Docs

Complete API reference, architecture guides, and tutorials in English and Chinese

API Docs Tutorials

Performance Comparison

Operation	OpenCV CPU	Mini-OpenCV GPU	Speedup
Gaussian Blur (4K)	45.2 ms	1.2 ms	37.7×
Sobel Edge (4K)	38.1 ms	0.9 ms	42.3×
Bilateral Filter (4K)	180.5 ms	4.8 ms	37.6×
Histogram Equalization (4K)	12.3 ms	0.3 ms	41.0×

Tested on RTX 4090 vs Intel i9-13900K, 3840×2160 images

Quick Start

cpp

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

// Create processor and load image
ImageProcessor processor;
GpuImage gpu = processor.loadFromHost(hostImage);

// Apply operations (all GPU-accelerated)
GpuImage blurred = processor.gaussianBlur(gpu, 5, 1.5f);
GpuImage edges = processor.sobelEdgeDetection(gpu);

// Download result
HostImage result = processor.downloadImage(edges);