import streamlit as st import torch import re from PIL import Image, ImageOps from transformers import TrOCRProcessor, VisionEncoderDecoderModel # ----------------------------- # CONFIG # ----------------------------- st.set_page_config(page_title="VISIO Hybrid OCR", layout="wide") TARGET_SIZE = (1024, 1024) # ----------------------------- # LOAD MODELS # ----------------------------- @st.cache_resource def load_models(): device = torch.device("cuda" if torch.cuda.is_available() else "cpu") proc_hw = TrOCRProcessor.from_pretrained("microsoft/trocr-large-handwritten") model_hw = VisionEncoderDecoderModel.from_pretrained( "microsoft/trocr-large-handwritten" ).to(device).eval() proc_pr = TrOCRProcessor.from_pretrained("microsoft/trocr-large-printed") model_pr = VisionEncoderDecoderModel.from_pretrained( "microsoft/trocr-large-printed" ).to(device).eval() proc_cv = TrOCRProcessor.from_pretrained("rijilraj77/trocr-curved-finetuned") model_cv = VisionEncoderDecoderModel.from_pretrained( "rijilraj77/trocr-curved-finetuned" ).to(device).eval() return { "handwritten": (proc_hw, model_hw), "printed": (proc_pr, model_pr), "curved": (proc_cv, model_cv), "device": device } # ----------------------------- # PREPROCESSING # ----------------------------- def preprocess_light(img: Image.Image) -> Image.Image: img = ImageOps.exif_transpose(img).convert("RGB") return img # ----------------------------- # OCR # ----------------------------- def ocr_trocr(img: Image.Image, proc, model, device) -> str: img = img.resize(TARGET_SIZE, Image.Resampling.LANCZOS) px = proc(images=img, return_tensors="pt").pixel_values.to(device) with torch.no_grad(): ids = model.generate(px, max_length=256) return proc.batch_decode(ids, skip_special_tokens=True)[0] # ----------------------------- # POST-PROCESSING # ----------------------------- def clean_text(text: str) -> str: text = re.sub(r"\s+", " ", text) text = text.replace("|", "I") return text.strip() # ----------------------------- # UI # ----------------------------- st.title("VISIO Hybrid OCR System") st.caption("Printed, Handwritten & Curved Text Recognition using TrOCR") models = load_models() mode = st.radio("Select OCR Mode", ["Printed", "Handwritten", "Curved"]) uploaded = st.file_uploader("Upload image", type=["png", "jpg", "jpeg"]) if uploaded: img = Image.open(uploaded) st.image(img, width=700) if st.button("Run OCR"): with st.spinner("Running OCR…"): img_prep = preprocess_light(img) proc, model = models[mode.lower()] device = models["device"] raw = ocr_trocr(img_prep, proc, model, device) st.session_state.raw_text = raw st.session_state.clean_text = clean_text(raw) st.success("OCR completed") if "raw_text" in st.session_state: st.subheader("Raw OCR Output") st.text_area("Raw", st.session_state.raw_text, height=200) st.subheader("Accurate Output (Post-processed)") st.text_area("Cleaned", st.session_state.clean_text, height=200) st.markdown("---") st.markdown( "**Notes:**\n" "• Upload clear word-level images for best results\n" "• Works for Printed, Handwritten and Curved text\n" "• Crop tightly around text for higher accuracy\n" )