PalOCR Model

Introduction

PalOCR is a CRNN ('None-VGG-BiLSTM-CTC') model trained base from EasyOCR guideline with solely purpose of getting a better score of openthaigpt/thai-ocr-evaluation datasets due to limitation of author hardware.

Training Dataset

Generated images of openthaigpt/thai-ocr-evaluation datasets using TextRecognitionDataGenerator. Which can be found at palocr-datasets

How to Use

Here’s how to use this model with EasyOCR: Please download, extract and place palocr.py, palocr.yaml in the user_network_directory (default = ~/.EasyOCR/user_network) and place palocr.pth in model directory (default = ~/.EasyOCR/model) Once you place all 3 files in their respective places you can use this code to run model.

import easyocr
reader = easyocr.Reader(["th", "en"], gpu=True, recog_network="palocr")
result = reader.readtext('text.jpg')

Model Performance Comparison

This section details the performance comparison between the open-source ThaiTrOCR model and other widely-used OCR systems, namely EasyOCR and Tesseract. The table below highlights their respective performance across various document types based on the average Character Error Rate (CER).

Category	EasyOCR	PalOCR	Tesseract
real_document	0.220217	0.960289	0.915707
scene_text	0.35865	1.0211	2.408704
handwritten	0.409302	1.01395	1.032375
document	0.0871795	0.946154	0.761595
document_enth	0.275449	0.916168	1.061107

Disclaimer: While this model is train on generated images of evaluation datasets, It was train on roughly 1,000 of generated images.

Key Insights

• Character Error Rate (CER): This metric evaluates the percentage of characters that were incorrectly predicted by the model. A lower CER indicates better performance. As shown in the table, ThaiTrOCR consistently outperforms EasyOCR and Tesseract across all document types, with a significantly lower average CER, making it the most accurate model in the comparison.
• Tesseract Limitation: It’s important to note that Tesseract only supports single-language input at a time in this comparison. For the purposes of this benchmark, it was tested using only the Thai language setting, which might have contributed to its higher CER values.
• The evaluation dataset is sourced from the openthaigpt/thai-ocr-evaluation.

Authors

• Vorakan Sumethsenee ([email protected])