thai-trocr

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image
import requests

# Load processor and model
processor = TrOCRProcessor.from_pretrained('openthaigpt/thai-trocr')
model = VisionEncoderDecoderModel.from_pretrained('openthaigpt/thai-trocr')

# Load an image
url = 'your_image_url_here'
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")

# Process and generate text
pixel_values = processor(images=image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(generated_text)

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).

Disclaimer: The test dataset at https://huggingface.co/datasets/openthaigpt/thai-ocr-evaluation includes only 104 images, which may limit the generalizability of these results. We are increasing the number of the test dataset.

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.
• Model Performance: The ThaiTrOCR model is particularly effective with PDF documents (both Thai-only and bilingual English-Thai texts), and shows substantial improvement over competing models in reading scene text and handwritten content.
• 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.

Sponsors

Authors

• Suchut Sapsathien (suchut@outlook.com)
• Jillaphat Jaroenkantasima (autsadang41@gmail.com)