This repository offers a rapid inference API for demonstrating OCR tasks.
The ONNX model, PaddleOCR-512_onnx.onnx, is derived from the official ch_PP-OCRv3_rec_infer.tar. For further details, please consult the paddle2onnx documentation.
For comprehensive model specifications, kindly utilize the Netron application.
Sample testing data is provided in the ./icdar_data directory for demonstration purposes.
Please install the necessary packages using the following command:
pip3 install -r requirements.txt
characters.txt: A file listing the characters that the OCR model can recognize.icdar_data: A directory containing sample images for OCR testing.PaddleOCR-512_onnx.onnx: The ONNX model file for OCR.test_paddle.py: A Python script to execute the OCR pipeline, encompassing both preprocessing and postprocessing stages.res: A directory for storing the results, including preprocessed images, input tensors, and OCR text outcomes.eval_paddle_onnx.py: A script designed for evaluating the accuracy of the OCR model.test.txt: A text file conatins evaluation inputs and its ground truth.requirements.txt: A text file conatins necessary packages.
Execute the following command for inference:
python3 test_paddle.py --input_tensor_save --opt_tensor_save --preprocess_img_save
Options:
--input_tensor_save: Saves the pre-processed input tensors.--opt_tensor_save: Saves the output tensors generated by the OCR model.--preprocess_img_save: Saves the pre-processed images for inspection.
Inspect the res directory for the results post-execution, which include preprocessed images, input and output tensors, and the OCR text outcomes.
The OCR procedure can be tailored by adjusting the arguments provided to test_paddle.py. For instance, modifications can be made to the input image directory, model file, and output directory. For a comprehensive list of options, please refer to the argument parser section within the script.
To evaluate the model, run:
python3 ./eval_paddle_onnx.py
This script calculates and displays the accuracy and normalized edit distance between the OCR model's predictions and the actual ground truth.
For assessing the model's efficacy, various test datasets under different scenarios are available in the ./data folder, including Chinese, English, and date/time samples.
Additional examples of preprocessed images are accessible in ./res/pre_res/.
The output text files document the predictions and their associated probabilities:
Predictions for ./icdar_data/12.png:
- Predicted text: 2022-05-09 星期一 01:38:58
- Prediction probability: 0.8843010663986206
Predictions for ./icdar_data/15.png:
- Predicted text: DS577杭州光大仓大件区505152
- Prediction probability: 0.9113200306892395
Predictions for ./icdar_data/word_10.png:
- Predicted text: PAIN
- Prediction probability: 0.9839118719100952
For interest in this project or any inquiries, please do not hesitate to contact me.