app.py

import gradio as gr
from PIL import Image
import torch
from transformers import AutoFeatureExtractor, AutoTokenizer, TrOCRProcessor, VisionEncoderDecoderModel
import os

# def sauvola_thresholding(grayImage_, window_size=15):
#     """
#     Sauvola thresholds are local thresholding techniques that are 
#     useful for images where the background is not uniform, especially for text recognition.
    
#     grayImage_ --- Input image should be in 2-Dimension Gray Scale format.
#     window_size --- It represents the filter window size.
#     """
#     # Assert the input conditions
#     assert len(grayImage_.shape) == 2, "Input image must be a 2-dimensional gray scale image."
#     assert isinstance(window_size, int) and window_size > 0, "Window size must be a positive integer."
    
#     thresh_sauvolavalue = threshold_sauvola(grayImage_, window_size=window_size)
#     thresholdImage_ = (grayImage_ > thresh_sauvolavalue)
    
#     return thresholdImage_

class OCRModel:
    def __init__(self, encoder_model, decoder_model, trained_model_path):
        # Load processor and model
        self.feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_model)
        self.decoder_tokenizer = AutoTokenizer.from_pretrained(decoder_model)
        self.processor = TrOCRProcessor(feature_extractor=self.feature_extractor, tokenizer=self.decoder_tokenizer)
        self.model = VisionEncoderDecoderModel.from_pretrained(trained_model_path)
        
        # Configure model settings
        self.model.config.decoder_start_token_id = self.processor.tokenizer.cls_token_id
        self.model.config.pad_token_id = self.processor.tokenizer.pad_token_id
        self.model.config.vocab_size = self.model.config.decoder.vocab_size
        self.model.config.eos_token_id = self.processor.tokenizer.sep_token_id
        self.model.config.max_length = 64
        self.model.config.early_stopping = True
        self.model.config.no_repeat_ngram_size = 3
        self.model.config.length_penalty = 2.0
        self.model.config.num_beams = 4
        
    def read_and_show(self, image_path):
        """
        Reads an image from the provided path and converts it to RGB.
        :param image_path: String, path to the input image.
        :return: PIL Image object
        """
        image = Image.open(image_path).convert('RGB')
        return image

    def ocr(self, image):
        """
        Performs OCR on the given image.
        :param image: PIL Image object.
        :return: Extracted text from the image.
        """
        # Preprocess the image
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        pixel_values = self.processor(image, return_tensors='pt').pixel_values.to(device)

        # Generate text
        generated_ids = self.model.generate(pixel_values)
        generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
        return generated_text

# Initialize the OCR model
ocr_model = OCRModel(encoder_model="google/vit-base-patch16-224-in21k",
                     decoder_model="surajp/RoBERTa-hindi-guj-san",
                     trained_model_path="./model/") #'sabaridsnfuji/Tamil_Offline_Handwritten_OCR')#"./model/")


    
def main(image_path):
# Process the image and extract text
    image = ocr_model.read_and_show(image_path)
    text = ocr_model.ocr(image)
    
    return image, text

# Gradio Interface function
def gradio_interface(image):
    # Save the uploaded image locally
    image_path = "uploaded_image.png"
    image.save(image_path)
    
    # Call the main function to process the image and get the result
    processed_image, result_text = main(image_path)
    
    return processed_image, result_text

# Sample images for demonstration (make sure these image paths exist)
sample_images = [
    "./sample/16.jpg",  # replace with actual image paths
    "./sample/20.jpg",  # replace with actual image paths
    "./sample/21.jpg",  # replace with actual image paths
    "./sample/31.jpg",  # replace with actual image paths
    "./sample/35.jpg",  # replace with actual image paths
]

# Ensure sample images directory exists
os.makedirs("samples", exist_ok=True)
# Save some dummy sample images if they don't exist (you should replace these with actual images)
for i, sample in enumerate(sample_images):
    if not os.path.exists(sample):
        img = Image.new("RGB", (224, 224), color=(i * 50, i * 50, i * 50))
        img.save(sample)

# Gradio UI setup with examples
gr_interface = gr.Interface(
    fn=gradio_interface, 
    inputs=gr.Image(type="pil"),  # Updated to gr.Image
    outputs=[gr.Image(type="pil"), gr.Textbox()],  # Updated to gr.Image and gr.Textbox
    title="Hindi Handwritten OCR Recognition", 
    description="Upload a cropped image containing a word, or use the sample images below to recognize the text. This is a word recognition model. Currently, text detection is not supported.",
    examples=sample_images  # Add the examples here
)

# Launch the Gradio interface
if __name__ == "__main__":
    gr_interface.launch()