English | 简体中文

Introduction

Though PaddleOCR provides support for TensorRT, it is difficult to decouple. This project provides simple code and demonstrates how to use the TensorRT C++ API and ONNX to deploy PaddleOCR text recognition model.

Enviroment

• Ubuntu 18.04
• CUDA 10.2
• cuDNN 8.4
• OpenCV 3.4.15
• TensorRT 8.4.1.5

Prerequisites

• Convert the trained model to the inference model.
• Convert the inference model to ONNX model

Details can be found in the PaddleOCR official document.

Doc-Text Recognition Model Inference

Doc-Paddle2ONNX (Chinese)

The model used in this tutorial are ch_PP-OCRv2_rec and ch_PP-OCRv3_rec from PaddleOCR model list. Download the corresponding inference model and convert it to ONNX model.

Usage

• Specify your path in CMakeLists.txt line 14-15.

# TODO: Specify the path to TensorRT root dir
set(TensorRT_DIR "/usr/yyx/tensorrt/TensorRT/")

• Modify the parameter in main.cpp line 14-23 if needed.

// TODO: Specify your precision here.
options.FP16 = false;
// TODO: Specify your input dimension here.
options.inputDimension = {3,48,320}; // Modify to {3,32,320} when using ppocrv2
// TODO: Specify your character_dict here.
std::string label_path = "../data/ppocr_keys_v1.txt";
// TODO: Specify your test image here.
const std::string inputImage = "../data/word_2.png";
// TODO: Specify your model here.
const std::string onnxModelpath = "../data/modelv3.onnx"; // Modify to "../data/modelv2.onnx" when using ppocrv2

• Building the library

mkdir build
cd build
cmake ..
make
./demo

Results

PPOCRv2

The result of ch_PP-OCRv2_rec ONNX model on data/word_2.png:

yourself        score: 0.95626300573349

The result of ch_PP-OCRv2_rec using tools/infer_rec.py in PaddleOCR on data/word_2.png:

{"Student": {"label": "yourself", "score": 0.9562630653381348}
 "Teacher": {"label": "yourself", "score": 0.9850824475288391}}

PPOCRv3

The result of ch_PP-OCRv3_rec ONNX model on data/word_2.png:

yourself        score: 0.9922693371772766

The result of ch_PP-OCRv3_rec using tools/infer_rec.py in PaddleOCR on data/word_2.png:

{"Student": {"label": "yourself", "score": 0.9922693371772766}
"Teacher": {"label": "yourself", "score": 0.9903509020805359}}

Thanks to

• tensorrt-cpp-api for creating a easy-to-use TensorRT C++ API Tutorial.

• PaddleOCRv2_TensorRT for creating some C++ implemention of PaddleOCR preprocess and postprocess method.

• PaddleOCR for creating awesome and practical OCR tools that help users train better models and apply them into practice.