TrOCR Finetuned for Dhivehi Text Recognition

A TrOCR model finetuned for Dhivehi (Divehi/Maldivian) text recognition using DeiT base encoder and BERT decoder.

Model Details

• Base models:
  • Encoder: facebook/deit-base-distilled-patch16-384
  • Decoder: alakxender/bert-base-dv
• Training data: 10k samples with 90/10 train/test split
• Input size: 384x384 pixels
• Beam search parameters:
  • max_length: 64
  • num_beams: 4
  • early_stopping: True
  • length_penalty: 2.0
  • no_repeat_ngram_size: 3

Training

The model was trained with:

• 7 epochs
• Batch size: 8
• Learning rate: 4e-5
• FP16 mixed precision
• Training augmentations:
  • Elastic transform (α=8.0, σ=5.0)
  • Gaussian blur (kernel size=(5,9), σ=(0.1,5))
  • Resize (384x384)
  • Normalization ([0.5,0.5,0.5], [0.5,0.5,0.5])

Usage

from PIL import Image
import torch
from torchvision import transforms
from transformers import (
   DeiTImageProcessor,
   TrOCRProcessor, 
   VisionEncoderDecoderModel,
   AutoTokenizer
)

class OCRPredictor:
   def __init__(self, model_name="alakxender/trocr-dv-diet-base-bert"):
       self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
       self.model = self._load_model(model_name)
       self.processor = self._load_processor()
       self.transform = self._get_transforms()
       
   def _load_model(self, model_name):
       model = VisionEncoderDecoderModel.from_pretrained(model_name)
       return model.to(self.device)
   
   def _load_processor(self):
       tokenizer = AutoTokenizer.from_pretrained("alakxender/trocr-dv-diet-base-bert")
       image_processor = DeiTImageProcessor.from_pretrained("facebook/deit-base-distilled-patch16-384")
       return TrOCRProcessor(image_processor=image_processor, tokenizer=tokenizer)
       
   def _get_transforms(self):
       return transforms.Compose([
           transforms.Resize((384, 384)),
           transforms.ToTensor(),
           transforms.Normalize([0.5] * 3, [0.5] * 3)
       ])
       
   def predict(self, image_path):
       image = Image.open(image_path).convert("RGB")
       pixel_values = self.transform(image).unsqueeze(0).to(self.device)
       
       outputs = self.model.generate(
           pixel_values,
           max_length=64,
           num_beams=4,
           early_stopping=True,
           length_penalty=2.0,
           no_repeat_ngram_size=3
       )
       
       return self.processor.decode(outputs[0], skip_special_tokens=True)
       
# Usage
predictor = OCRPredictor()
text = predictor.predict("ocr2.png")
print(text) # ތިން މިނިސްޓްރީއެއް ހިންގާ މ.ގްރީން ބިލްޑިންގުގައި މިދިޔަ ބުރާސްފަތި ދުވަހު ހިނގި ބޮޑު އަލިފާނުގެ.

Evaluation Results

[
  {
    "file_name": "data/images/DV01-04/DV01-04_140.jpg",
    "predicted_text": "ޤާނޫނުގެ 42 ވަނަ މާއްދާގައި ލާޒިމްކުރާ މި ރިޕޯޓު ތައްޔާރުކޮށް ފޮނުވުމުގެ ޒިންމާއަކީ ޤާނޫނުން އިދާރާގެ އިންފޮމޭޝަން އޮފިސަރު ކުރައްވަންޖެހޭ ކަމެކެވެ .",
    "true_text": "ޤާނޫނުގެ 42 ވަނަ މާއްދާގައި ލާޒިމްކުރާ މި ރިޕޯޓު ތައްޔާރުކޮށް ފޮނުވުމުގެ ޒިންމާއަކީ ޤާނޫނުން އިދާރާގެ އިންފޮމޭޝަން އޮފިސަރު ކުރައްވަންޖެހޭ ކަމެކެވެ."
  }
]