temp_app.py

import streamlit as st
from PIL import Image
from surya.ocr import run_ocr
from surya.model.detection.model import load_model as load_det_model, load_processor as load_det_processor
from surya.model.recognition.model import load_model as load_rec_model
from surya.model.recognition.processor import load_processor as load_rec_processor
import re
from transformers import AutoModel, AutoTokenizer
import torch
import tempfile
import os

os.environ["CUDA_VISIBLE_DEVICES"] = ""

st.set_page_config(page_title="OCR Application", page_icon="🖼️", layout="wide")

# Force CPU if CUDA is not available
device = "cuda" if torch.cuda.is_available() else "cpu"

@st.cache_resource
def load_surya_models():
    det_processor, det_model = load_det_processor(), load_det_model()
    det_model.to(device)
    rec_model, rec_processor = load_rec_model(), load_rec_processor()
    rec_model.to(device)
    return det_processor, det_model, rec_model, rec_processor

@st.cache_resource
def load_got_ocr_model():
    tokenizer = AutoTokenizer.from_pretrained('ucaslcl/GOT-OCR2_0', trust_remote_code=True)
    model = AutoModel.from_pretrained('ucaslcl/GOT-OCR2_0', trust_remote_code=True, low_cpu_mem_usage=True, device_map=device, use_safetensors=True, pad_token_id=tokenizer.eos_token_id)
    model.eval().to(device)
    
    # Override .half() and .cuda() to ensure everything runs in float32 and on CPU
    torch.Tensor.half = lambda x: x.float()
    torch.Tensor.cuda = lambda x, **kwargs: x.cpu()

    return tokenizer, model

det_processor, det_model, rec_model, rec_processor = load_surya_models()
tokenizer, got_model = load_got_ocr_model()

st.title("OCR Application  (Aarish Shah Mohsin)")
st.write("Upload an image for OCR processing. Using GOT-OCR for English translations, Picked Surya OCR Model for English+Hindi Translations")

st.sidebar.header("Configuration")
model_choice = st.sidebar.selectbox("Select OCR Model:", ("For English + Hindi", "For English (GOT-OCR)"))

# Store the uploaded image and extracted text in session state
if 'uploaded_image' not in st.session_state:
    st.session_state.uploaded_image = None
if 'extracted_text' not in st.session_state:
    st.session_state.extracted_text = ""

uploaded_file = st.sidebar.file_uploader("Choose an image...", type=["png", "jpg", "jpeg"])

# Update the session state if a new file is uploaded
if uploaded_file is not None:
    st.session_state.uploaded_image = uploaded_file

predict_button = st.sidebar.button("Predict", key="predict")

col1, col2 = st.columns([2, 1])

# Display the image preview if it's already uploaded
if st.session_state.uploaded_image:
    image = Image.open(st.session_state.uploaded_image)

    with col1:
        # Display a smaller preview of the uploaded image (set width to 300px)
        col1.image(image, caption='Uploaded Image', use_column_width=False, width=300)

# Handle predictions
if predict_button and st.session_state.uploaded_image:
    with st.spinner("Processing..."):
        # Save the uploaded file temporarily
        with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
            temp_file.write(st.session_state.uploaded_image.getvalue())
            temp_file_path = temp_file.name

        image = Image.open(temp_file_path)
        image = image.convert("RGB")  

        if model_choice == "For English + Hindi":
            langs = ["en", "hi"]
            predictions = run_ocr([image], [langs], det_model, det_processor, rec_model, rec_processor)
            text_list = re.findall(r"text='(.*?)'", str(predictions[0]))
            extracted_text = ' '.join(text_list)

            st.session_state.extracted_text = extracted_text  # Save extracted text in session state

        elif model_choice == "For English (GOT-OCR)":
            image_file = temp_file_path
            res = got_model.chat(tokenizer, image_file, ocr_type='ocr')

            st.session_state.extracted_text = res  # Save extracted text in session state

        # Delete the temporary file after processing
        if os.path.exists(temp_file_path):
            os.remove(temp_file_path)

# Search functionality
if st.session_state.extracted_text:
    search_query = st.text_input("Search in extracted text:", key="search_query", placeholder="Type to search...")

    # Create a pattern to find the search query in a case-insensitive way
    if search_query:
        pattern = re.compile(re.escape(search_query), re.IGNORECASE)
        highlighted_text = st.session_state.extracted_text
        
        # Replace matching text with highlighted version (bright green)
        highlighted_text = pattern.sub(lambda m: f"<span style='background-color: limegreen;'>{m.group(0)}</span>", highlighted_text)

        st.markdown("### Highlighted Search Results:")
        st.markdown(highlighted_text, unsafe_allow_html=True)
    else:
        # If no search query, show the original extracted text
        st.markdown("### Extracted Text:")
        st.markdown(st.session_state.extracted_text, unsafe_allow_html=True)