Update app.py

app.py

@@ -1,77 +1,77 @@
-#!/usr/bin/env python3
-"""
-DocFusion — Gradio Web UI
-Rihal CodeStacker 2026
-"""
+
 
 import os
-os.system("apt-get install -y tesseract-ocr")
-os.system("pip install pytesseract -q")
+import subprocess
+
+
+libs = ["easyocr", "transformers", "torchvision", "gradio", "huggingface_hub"]
+for lib in libs:
+    subprocess.run(["pip", "install", lib, "-q"])
 
 import re
 import json
 import torch
 import numpy as np
+import easyocr
+import gradio as gr
 from PIL import Image, ImageDraw
 from torchvision import transforms, models
 from transformers import LayoutLMForTokenClassification, BertTokenizerFast
 from huggingface_hub import hf_hub_download
-import gradio as gr
+
 
 DEVICE   = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 LABELS   = ["O", "B-VENDOR", "I-VENDOR", "B-DATE", "I-DATE", "B-TOTAL", "I-TOTAL"]
 ID2LABEL = {i: x for i, x in enumerate(LABELS)}
 LABEL2ID = {x: i for i, x in enumerate(LABELS)}
 
-print("Loading models...")
-tokenizer        = BertTokenizerFast.from_pretrained("microsoft/layoutlm-base-uncased")
+print(f"Status: Loading models on {DEVICE}...")
+
+
+reader = easyocr.Reader(['en'], gpu=torch.cuda.is_available())
+
+
+tokenizer = BertTokenizerFast.from_pretrained("microsoft/layoutlm-base-uncased")
 extraction_model = LayoutLMForTokenClassification.from_pretrained(
     "Zakariya007/docfusion-v1",
     num_labels=len(LABELS), id2label=ID2LABEL, label2id=LABEL2ID,
-)
-extraction_model = extraction_model.to(DEVICE)
+).to(DEVICE)
 extraction_model.eval()
 
+
 forgery_model = models.efficientnet_b0(weights=None)
 forgery_model.classifier[1] = torch.nn.Linear(1280, 2)
-weights_path = hf_hub_download(
-    repo_id="Zakariya007/docfusion-v2",
-    filename="efficientnet_best.pth"
-)
+weights_path = hf_hub_download(repo_id="Zakariya007/docfusion-v2", filename="efficientnet_best.pth")
 forgery_model.load_state_dict(torch.load(weights_path, map_location=DEVICE))
 forgery_model = forgery_model.to(DEVICE)
 forgery_model.eval()
-print("✅ Models loaded!")
+
+print("All systems ready!")
 
 def extract_fields(image):
     try:
-        import pytesseract
-        ocr_text = pytesseract.image_to_string(image)
-
-        # Regex-based extraction
-        date_match = re.search(
-            r'\d{1,2}[\/\-\.]\d{1,2}[\/\-\.]\d{2,4}',
-            ocr_text
-        )
+        img_np = np.array(image)
+        results = reader.readtext(img_np)
+
+        full_text = " ".join([res[1] for res in results])
+        lines = [res[1].strip() for res in results if len(res[1].strip()) > 2]
+
+
+        date_match = re.search(r'\d{1,2}[\/\-\.]\d{1,2}[\/\-\.]\d{2,4}', full_text)
         date = date_match.group(0) if date_match else None
 
-        total_match = re.search(
-            r'(?:TOTAL|AMOUNT|JUMLAH)[^\d]*(\d+[\.,]\d{2})',
-            ocr_text, re.IGNORECASE
-        )
+        total_match = re.search(r'(?:TOTAL|AMOUNT|NET|DUE|CASH|SUBTOTAL)[^\d]*([\d,]+\.\d{2})', full_text, re.IGNORECASE)
         total = total_match.group(1) if total_match else None
 
-        lines  = [l.strip() for l in ocr_text.split('\n') if len(l.strip()) > 3]
         vendor = lines[0] if lines else None
 
         return vendor, date, total
-
     except Exception as e:
         print(f"Extraction error: {e}")
         return None, None, None
 
-
-def detect_forgery(image, vendor, date, total):
+def detect_forgery_pure_model(image):
+    """Detects forgery using only the neural network output."""
     transform = transforms.Compose([
         transforms.Resize((224, 224)),
         transforms.ToTensor(),
@@ -79,106 +79,70 @@ def detect_forgery(image, vendor, date, total):
     ])
     tensor = transform(image).unsqueeze(0).to(DEVICE)
     with torch.no_grad():
-        output      = forgery_model(tensor)
-        probs       = torch.softmax(output, dim=1)
+        output = forgery_model(tensor)
+        probs = torch.softmax(output, dim=1)
         forged_prob = probs[0][1].item()
-        visual_flag = forged_prob > 0.5
-
-    rule_flags = []
-    if not vendor: rule_flags.append("Missing vendor")
-    if not date:   rule_flags.append("Missing date")
-    if not total:  rule_flags.append("Missing total")
-
-    if total:
-        try:
-            total_val = float(re.sub(r"[^\d.]", "", total))
-            if total_val > 10000: rule_flags.append("Abnormally high total")
-            if total_val <= 0:    rule_flags.append("Invalid total amount")
-        except Exception:
-            rule_flags.append("Unparseable total")
-
-    if date:
-        date_pattern = re.compile(
-            r"\d{1,2}[\/\-\.]\d{1,2}[\/\-\.]\d{2,4}|\d{4}[\/\-\.]\d{2}[\/\-\.]\d{2}"
-        )
-        if not date_pattern.search(date):
-            rule_flags.append("Invalid date format")
-
-    rule_flag = len(rule_flags) >= 2
-    is_forged = 1 if (visual_flag or rule_flag) else 0
-    return is_forged, forged_prob, rule_flags
-
-
-def annotate_image(image, is_forged):
-    annotated = image.copy()
-    draw      = ImageDraw.Draw(annotated)
-    w, h      = image.size
-    if is_forged:
-        draw.rectangle([0, 0, w-1, h-1], outline="#FF0000", width=6)
-        draw.text((10, 10), "FORGED", fill="#FF0000")
-    else:
-        draw.rectangle([0, 0, w-1, h-1], outline="#00AA00", width=6)
-        draw.text((10, 10), "GENUINE", fill="#00AA00")
-    return annotated
 
 
+    is_forged = 1 if forged_prob > 0.5 else 0
+    return is_forged, forged_prob
+
 def process_receipt(image):
-    if image is None:
-        return None, "Please upload an image.", "", "", "", ""
-
-    pil_image = Image.fromarray(image).convert("RGB")
-    vendor, date, total        = extract_fields(pil_image)
-    is_forged, forged_prob, rule_flags = detect_forgery(pil_image, vendor, date, total)
-    annotated                  = annotate_image(pil_image, is_forged)
-
-    if is_forged:
-        status = f"FORGED (confidence: {forged_prob:.1%})"
-        if rule_flags:
-            status += "\nFlags: " + ", ".join(rule_flags)
-    else:
-        status = f"GENUINE (forged probability: {forged_prob:.1%})"
-
-    return (
-        np.array(annotated),
-        status,
-        vendor or "Not detected",
-        date   or "Not detected",
-        total  or "Not detected",
-        json.dumps({
-            "vendor":    vendor,
-            "date":      date,
-            "total":     total,
-            "is_forged": is_forged
-        }, indent=2),
-    )
+    if image is None: return None, "No image uploaded", "", "", "", ""
+
+    pil_img = Image.fromarray(image).convert("RGB")
+
+
+    vendor, date, total = extract_fields(pil_img)
+
 
+    is_forged, prob = detect_forgery_pure_model(pil_img)
 
-with gr.Blocks(title="DocFusion") as demo:
-    gr.Markdown("# DocFusion — Intelligent Document Processing")
-    gr.Markdown("Upload a scanned receipt to extract fields and detect forgery.")
+
+    annotated = pil_img.copy()
+    draw = ImageDraw.Draw(annotated)
+    color = "#FF4B4B" if is_forged else "#24A148"
+    draw.rectangle([0, 0, pil_img.size[0]-1, pil_img.size[1]-1], outline=color, width=12)
+
+    status = f"Result: {'SUSPECTED FORGERY' if is_forged else 'LIKELY GENUINE'}"
+    confidence_str = f"Model Confidence: {prob if is_forged else (1-prob):.1%}"
+    full_status = f"{status}\n{confidence_str}"
+
+    res_json = json.dumps({
+        "vendor": vendor,
+        "date": date,
+        "total": total,
+        "forgery_score": round(prob, 4),
+        "is_forged": bool(is_forged)
+    }, indent=2)
+
+    return np.array(annotated), full_status, vendor or "N/A", date or "N/A", total or "N/A", res_json
+
+
+with gr.Blocks(theme=gr.themes.Default()) as demo:
+    gr.Markdown("# 📑 DocFusion: Receipt Intelligence (V2)")
+    gr.Markdown("Visual Forgery Detection + Deep Learning OCR")
 
     with gr.Row():
         with gr.Column():
-            input_image = gr.Image(label="Upload Receipt", type="numpy")
-            submit_btn  = gr.Button("Analyze Receipt", variant="primary")
+            in_img = gr.Image(label="Upload Receipt Scan", type="numpy")
+            btn = gr.Button("Analyze Receipt", variant="primary")
         with gr.Column():
-            output_image   = gr.Image(label="Annotated Receipt")
-            forgery_status = gr.Textbox(label="Forgery Status", lines=4)
+            out_img = gr.Image(label="Visual Analysis")
+            out_stat = gr.Textbox(label="Forgery Detection Status", lines=2)
 
     with gr.Row():
-        vendor_out = gr.Textbox(label="Vendor")
-        date_out   = gr.Textbox(label="Date")
-        total_out  = gr.Textbox(label="Total")
+        v_out = gr.Textbox(label="Vendor")
+        d_out = gr.Textbox(label="Date")
+        t_out = gr.Textbox(label="Total Amount")
 
-    json_out = gr.Code(label="JSON Output", language="json")
+    js_out = gr.Code(label="Metadata Output (JSON)", language="json")
 
-    submit_btn.click(
-        fn      = process_receipt,
-        inputs  = [input_image],
-        outputs = [output_image, forgery_status, vendor_out, date_out, total_out, json_out],
+    btn.click(
+        process_receipt,
+        inputs=[in_img],
+        outputs=[out_img, out_stat, v_out, d_out, t_out, js_out]
     )
 
-    gr.Markdown("**Models:** LayoutLM v1 (extraction) + EfficientNet-B0 (forgery)")
-
 if __name__ == "__main__":
-    demo.launch(share=True)
+    demo.launch(debug=True)