milkv-PaddleOCR-ncnn

This repository provides a C++ implementation of PaddleOCR optimized for RISC-V architecture, using the ncnn inference framework for efficient OCR processing on Milk-V platforms.

System Requirements

Dependencies

• OpenCV
• ncnn inference framework
• risc-toolchain

Installation Guide

1. Download Dependencies

1.1 Download and Build OpenCV

For x86 platform:

# Run OpenCV build script
./tools/build_opencv.sh

For RISC-V platform:

# Run RISC-V OpenCV build script
./tools/build_opencv_risc.sh

After compilation, OpenCV will be installed in:

• x86: lib/opencv/opencv4/
• RISC-V: lib/opencv/opencv4_riscv/

1.2 Download and Build ncnn

For x86 platform:

# Enter ncnn directory
cd lib/ncnn

# Run build script
./build.sh

# After completion, lib/cmake/ncnn directory will be generated
# You need to change lib PATH

For RISC-V platform:

# Run RISC-V ncnn build script
./tool/build_ncnn_risc.sh

After compilation, ncnn will be installed in:

• x86: lib/ncnn/lib/cmake/ncnn/
• RISC-V: lib/ncnn/ncnn_riscv/lib/cmake/ncnn/

2. Download Model Files

Model files should be placed in the models/ directory, including:

models/
├── det.bin                    # Detection model binary file
├── det.param                  # Detection model parameter file
├── cls.bin                    # Classification model binary file
├── cls.param                  # Classification model parameter file
├── rec.bin                    # Recognition model binary file
├── rec.param                  # Recognition model parameter file
└── keys.txt                   # Character dictionary file

Recommended Models:

• PP-OCRv3 series models
• PP-OCRv5 series models (mobile/server versions)

3. Build Project

3.1 x86 Platform Build

# Create build directory
mkdir -p build
cd build

# Configure CMake
cmake ..

# Build
make -j$(nproc)

3.2 RISC-V Platform Build

# Use provided build script
./tools/build.sh

Or manual build:

# Create build directory
mkdir -p build
cd build

# Configure CMake (enable RISC-V cross-compilation)
cmake .. -DBUILD_FOR_RISCV=ON

# Build
make -j$(nproc)

Usage

Basic Usage

# Syntax
./build/main <config_file> <image_path>

# Example
./build/main config.json images/test.jpg

Configuration File

Edit config.json to adjust OCR parameters:

{
    "save": true,
    "det": {
        "infer_threads": -1,          // Inference thread count (-1 for auto)
        "model_path": "../models/det", // Detection model path
        "padding": 0,                 // Padding size
        "max_side_len": 768,          // Maximum side length
        "box_thres": 0.5,             // Box threshold
        "bitmap_thres": 0.3,          // Binarization threshold
        "unclip_ratio": 2.0,          // Expansion ratio
        "fp16": false                 // Whether to use FP16
    },
    "cls": {
        "infer_threads": 1,           // Inference thread count
        "reco_threads": -1,           // Recognition thread count
        "model_path": "../models/cls", // Classification model path
        "enable": true,               // Whether to enable classification
        "most_angle": true,           // Whether to use most common angle
        "fp16": false                 // Whether to use FP16
    },
    "rec": {
        "infer_threads": 1,           // Inference thread count
        "reco_threads": -1,           // Recognition thread count
        "model_path": "../models/rec", // Recognition model path
        "keys_path": "../models/keys.txt", // Dictionary file path
        "fp16": false                 // Whether to use FP16
    }
}

Output Results

The program will output recognized text content, including:

• Text region coordinates
• Confidence scores
• Recognized text content

Directory Structure

PaddleOCR-ncnn-CPP/
├── src/                      # Source code
│   ├── main.cpp             # Main program entry
│   ├── ocr_engine.cpp       # OCR engine implementation
│   ├── db_net.cpp           # DB detection network
│   ├── angle_net.cpp        # Angle classification network
│   ├── crnn_net.cpp         # CRNN recognition network
│   ├── utils.cpp            # Utility functions
│   └── 3rdparty/           # Third-party libraries
├── models/                  # Model files directory
├── images/                  # Test images
├── lib/                     # Dependencies
│   ├── opencv/             # OpenCV library
│   └── ncnn/               # ncnn library
├── tools/                   # Build tools
│   ├── build.sh            # RISC-V build script
│   ├── build_opencv.sh     # OpenCV build script
│   ├── build_opencv_risc.sh # RISC-V OpenCV build script
│   └── build_ncnn_risc.sh  # RISC-V ncnn build script
├── config.json             # Configuration file
└── CMakeLists.txt          # CMake configuration

Performance Optimization

Thread Configuration

• infer_threads: Inference thread count, -1 means using all available cores
• reco_threads: Recognition thread count for parallel processing of multiple text regions

Model Selection

• Mobile Models: Smaller and faster, suitable for resource-constrained devices
• Server Models: Larger and more accurate, suitable for high-performance requirements

FP16 Acceleration

• Enable "fp16": true on supported hardware to improve inference speed

Troubleshooting

Common Issues

1. OpenCV not found

   Solution: Ensure OpenCV is correctly compiled and installed, check path settings in CMakeLists.txt
   

2. ncnn not found

   Solution: Ensure ncnn is correctly compiled, check ncnn_DIR path settings
   

3. Model loading failed

   Solution: Check model file paths, ensure both .bin and .param files exist
   

4. RISC-V cross-compilation failed

   Solution: Ensure RISC-V toolchain is correctly installed, check tools/toolchain/ directory
   

Debug Mode

Modify log level in src/main.cpp:

// Detailed output
plog::init(plog::debug, &console_appender);

// Basic output
plog::init(plog::info, &console_appender);

License

This project follows the license terms in the LICENSE file.

Acknowledgments

• Avafly/PaddleOCR-ncnn-CPP
• PaddleOCR
• ncnn
• OpenCV

Contributing

Issues and Pull Requests are welcome to improve this project!