PP-OCRv5 DEEPX Baseline

中文 README

🚀 PP-OCRv5 DEEPX benchmarking toolchain with NPU acceleration and comprehensive performance evaluation.

📈 Performance Results

Custom Dataset Overview

This project uses a diverse custom Chinese dataset for benchmarking. The dataset consists of various real-world scenarios including street signs, handwritten text, exam papers, textbooks, and newspapers, providing comprehensive coverage of different text recognition challenges with detailed annotations including text content and bounding box coordinates.

Test Configuration:

• Dataset: Custom Chinese document dataset (20 images)
• Data Format: PNG images with JSON annotations containing text content
• Model: DXNN-OCR v5 full pipeline (PP-OCRv5 → DEEPX NPU accelerated)
  • Text detection: PP-OCRv5 det → DXNN det_v5 (NPU accelerated)
  • Text classification: PP-OCRv5 cls → DXNN cls_v5 (NPU accelerated)
  • Text recognition: PP-OCRv5 rec → DXNN rec_v5 multi-ratio models (NPU accelerated)
• Hardware configuration:
  • Platform: Rockchip RK3588 IR88MX01 LP4X V10
  • NPU: DEEPX DX-M1 Accelerator Card
    • PCIe: Gen3 X4 interface [01:00:00]
    • Firmware: v2.1.0
  • CPU: ARM Cortex-A55 8-core @ 2.35GHz (8nm process)
  • Memory: 8GB LPDDR4X
  • Operating System: Ubuntu 20.04.6 LTS (Focal)
  • Runtime: DXRT v3.0.0 + RT driver v1.7.1 + PCIe driver v1.4.1

Benchmark Results:

NPU Model	Average Inference Time (ms)	Average FPS	Average CPS (chars/s)	Average Accuracy (%)
DEEPX DX-M1	1151.77	2.94	255.17	68.56

• Detailed Performance Results of PP-OCRv5 on DEEPX NPU

🛠️ Quick Start

⚡ One Simple Step to Start Your OCR Benchmark

One-Step Execution:

git clone https://github.com/Chris-godz/PP-OCRv5-DeepX-Baseline.git
cd PP-OCRv5-DeepX-Baseline
./startup.sh

📁 Project Structure

├── startup.sh              # One-click benchmark execution
├── scripts/
│   ├── dxnn_benchmark.py   # Main benchmark tool (NPU inference + performance testing)
│   ├── calculate_acc.py    # PP-OCRv5 compatible accuracy calculation
│   └── ocr_engine.py       # DXNN NPU engine interface
├── engine/
│   ├── model_files/v5/     # DXNN v5 NPU models (.dxnn format)
│   ├── draw_utils.py       # Visualization utilities
│   ├── utils.py           # Processing utilities
│   └── fonts/             # Chinese fonts (for visualization)
├── images/                 # Custom dataset (20 PNG images + labels.json)
│   ├── image_1.png ~ image_20.png  # Test images
│   └── labels.json         # Ground truth annotations
├── output/                # Test results output
│   ├── json/              # Detailed JSON results
│   ├── vis/               # Visualization images
│   ├── benchmark_summary.json
│   ├── benchmark_results.csv
│   └── DXNN-OCR_benchmark_report.md
└── logs/                  # Execution logs

Custom Dataset:

# Prepare your own images
mkdir -p images/custom
cp /path/to/your/images/* images/custom/

# Run benchmark
python scripts/dxnn_benchmark.py \
    --directory images/custom \
    --ground-truth custom_labels.json \
    --output output_custom \
    --runs 3

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

This project is forked and developed based on DEEPX-AI/DXNN-OCR project

• Thanks to DEEPX team for NPU runtime and foundational framework support
• Thanks to the PaddleOCR team for the excellent OCR framework