Update app.py

app.py

@@ -1,34 +1,35 @@
 import os
-
-# --- CẤU HÌNH HỆ THỐNG ---
-os.environ["FLAGS_use_mkldnn"] = "0"
-os.environ["FLAGS_enable_mkldnn"] = "0"
-os.environ["DN_ENABLE_MKLDNN"] = "0"
-os.environ["CPP_MIN_LOG_LEVEL"] = "3"
-
+import base64
+import io
+import json
 import logging
 import re
-import gradio as gr
-from paddleocr import PaddleOCR
-from PIL import Image, ImageDraw, ImageFont
+import cv2
 import numpy as np
 import requests
+from PIL import Image, ImageDraw, ImageFont
+import gradio as gr
+from paddleocr import PaddleOCR
 
-# Tắt log thừa
+# --- PHẦN 1: CẤU HÌNH & KHỞI TẠO PADDLEOCR (LOCAL ENGINE) ---
+os.environ["FLAGS_use_mkldnn"] = "0"
+os.environ["FLAGS_enable_mkldnn"] = "0"
+os.environ["CPP_MIN_LOG_LEVEL"] = "3"
 logging.getLogger("ppocr").setLevel(logging.WARNING)
 
-print("Đang khởi tạo PaddleOCR (Coordinate Sync Mode)...")
-
+print("🚀 Đang khởi tạo PaddleOCR Local...")
 try:
-    ocr = PaddleOCR(use_textline_orientation=True, use_doc_orientation_classify=False, 
-        use_doc_unwarping=False, lang='ch')
+    # Cấu hình OCR Local
+    ocr = PaddleOCR(use_textline_orientation=True, 
+                    use_doc_orientation_classify=False,
+                    use_doc_unwarping=False, 
+                    lang='ch') # Có thể đổi sang 'en' hoặc 'vi'
 except Exception as e:
-    print(f"Lỗi khởi tạo: {e}. Chuyển về chế độ mặc định.")
+    print(f"⚠️ Lỗi khởi tạo nâng cao: {e}. Dùng chế độ mặc định.")
     ocr = PaddleOCR(lang='ch')
+print("✅ Model đã sẵn sàng!")
 
-print("Model đã sẵn sàng!")
-
-# --- TẢI FONT ---
+# Tải Font để vẽ chữ (Từ Phần 1)
 def check_and_download_font():
     font_path = "./simfang.ttf"
     if not os.path.exists(font_path):
@@ -43,15 +44,19 @@ def check_and_download_font():
 
 FONT_PATH = check_and_download_font()
 
-# --- HÀM VẼ ĐA NĂNG ---
+# --- HELPER FUNCTIONS (HỖ TRỢ XỬ LÝ ẢNH & TEXT) ---
+
+def pil_to_base64_html(image):
+    """Chuyển đổi PIL Image thành thẻ HTML <img> base64"""
+    buffered = io.BytesIO()
+    image.save(buffered, format="JPEG")
+    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+    return f'<img src="data:image/jpeg;base64,{img_str}" alt="Result" style="width:100%; object-fit:contain;">'
+
 def universal_draw(image, raw_data, font_path):
+    """Hàm vẽ box lên ảnh (Từ Phần 1)"""
     if image is None: return image
-    
-    # Đảm bảo image là PIL
-    if isinstance(image, np.ndarray):
-        image = Image.fromarray(image)
-    
-    # Copy để vẽ
+    if isinstance(image, np.ndarray): image = Image.fromarray(image)
     canvas = image.copy()
     draw = ImageDraw.Draw(canvas)
     
@@ -61,161 +66,197 @@ def universal_draw(image, raw_data, font_path):
     except:
         font = ImageFont.load_default()
 
-    # Hàm parse box
-    def parse_box(b):
-        try:
-            if hasattr(b, 'tolist'): b = b.tolist()
-            if len(b) > 0 and isinstance(b[0], list): return [tuple(p) for p in b]
-            if len(b) == 4 and isinstance(b[0], (int, float)): 
-                 return [(b[0], b[1]), (b[2], b[1]), (b[2], b[3]), (b[0], b[3])]
-            return None
-        except: return None
-
-    items_to_draw = []
-
-    # Logic tìm box/text
-    # Ưu tiên cấu trúc PaddleX: rec_texts + dt_polys
-    processed = False
-    if isinstance(raw_data, list) and len(raw_data) > 0 and isinstance(raw_data[0], dict):
-        data_dict = raw_data[0]
-        texts = data_dict.get('rec_texts')
-        boxes = data_dict.get('dt_polys', data_dict.get('rec_polys', data_dict.get('dt_boxes')))
-        
-        if texts and boxes and isinstance(texts, list) and isinstance(boxes, list):
-            for i in range(min(len(texts), len(boxes))):
-                txt = texts[i]
-                box = parse_box(boxes[i])
-                if box and txt: items_to_draw.append((box, txt))
-            processed = True
-
-    # Fallback Logic
-    if not processed:
-        def hunt(data):
-            if isinstance(data, dict):
-                box = None; text = None
-                for k in ['points', 'box', 'dt_boxes', 'poly']:
-                    if k in data: box = parse_box(data[k]); break
-                for k in ['transcription', 'text', 'rec_text', 'label']:
-                    if k in data: text = data[k]; break
-                if box and text: items_to_draw.append((box, text)); return
-                for v in data.values(): hunt(v)
-            elif isinstance(data, (list, tuple)):
-                if len(data) == 2 and isinstance(data[0], list) and len(data[0]) == 4:
-                    box = parse_box(data[0])
-                    txt_obj = data[1]
-                    text = txt_obj[0] if isinstance(txt_obj, (list, tuple)) else txt_obj
-                    if box and isinstance(text, str): items_to_draw.append((box, text)); return
-                for item in data: hunt(item)
-        hunt(raw_data)
-
-    # Vẽ
-    for box, txt in items_to_draw:
-        try:
-            # Vẽ khung đỏ
-            draw.polygon(box, outline="red", width=3)
-            # Vẽ chữ
-            txt_x, txt_y = box[0]
-            if hasattr(draw, "textbbox"):
-                text_bbox = draw.textbbox((txt_x, txt_y), txt, font=font, anchor="lb")
-                draw.rectangle(text_bbox, fill="red")
-                draw.text((txt_x, txt_y), txt, fill="white", font=font, anchor="lb")
-            else:
-                draw.text((txt_x, txt_y - font_size), txt, fill="white", font=font)
-        except: continue
+    boxes = [line[0] for line in raw_data[0]] if raw_data and raw_data[0] else []
+    txts = [line[1][0] for line in raw_data[0]] if raw_data and raw_data[0] else []
+    scores = [line[1][1] for line in raw_data[0]] if raw_data and raw_data[0] else []
 
+    for box, txt in zip(boxes, txts):
+        box = [tuple(p) for p in box]
+        draw.polygon(box, outline="red", width=3)
+        # Vẽ nền chữ
+        if hasattr(draw, "textbbox"):
+            text_bbox = draw.textbbox(box[0], txt, font=font, anchor="lb")
+            draw.rectangle(text_bbox, fill="red")
+            draw.text(box[0], txt, fill="white", font=font, anchor="lb")
+        else:
+            draw.text((box[0][0], box[0][1] - font_size), txt, fill="white", font=font)
     return canvas
 
-# --- HÀM XỬ LÝ TEXT ---
-def deep_extract_text(data):
-    found_texts = []
-    if isinstance(data, str):
-        if len(data.strip()) > 0: return [data]
-        return []
-    if isinstance(data, (list, tuple)):
-        for item in data: found_texts.extend(deep_extract_text(item))
-    elif isinstance(data, dict):
-        for val in data.values(): found_texts.extend(deep_extract_text(val))
-    elif hasattr(data, '__dict__'): found_texts.extend(deep_extract_text(data.__dict__))
-    return found_texts
-
-def clean_text_result(text_list):
-    cleaned = []
-    block_list = ['min', 'max', 'general', 'header', 'footer', 'structure']
-    for t in text_list:
-        t = t.strip()
-        if len(t) < 2 and not any(u'\u4e00' <= c <= u'\u9fff' for c in t): continue
-        if t.lower().endswith(('.ttf', '.json', '.pdparams', '.yml', '.log')): continue
-        if t.lower() in block_list: continue
-        if not re.search(r'[\w\u4e00-\u9fff]', t): continue
-        cleaned.append(t)
-    return cleaned
-
-# --- MAIN PREDICT ---
-def predict(image):
-    if image is None: return None, "Chưa có ảnh.", "No Data"
+# --- HÀM XỬ LÝ CHÍNH (LOGIC CẦU NỐI) ---
+# Hàm này nhận input từ UI Phần 2, chạy Logic Phần 1, trả về format UI Phần 2
+
+def local_inference(image_path, mode="Document"):
+    if not image_path:
+        return "Please upload an image.", "", ""
 
     try:
-        # Chuẩn bị ảnh đầu vào
-        original_pil = image.copy() if isinstance(image, Image.Image) else Image.fromarray(image).copy()
-        image_np = np.array(image)
-        
-        # 1. OCR
-        raw_result = ocr.ocr(image_np)
-        
-        # 2. XỬ LÝ ẢNH ĐỂ VẼ (KEY FIX: Lấy ảnh từ Preprocessor nếu có)
-        target_image_for_drawing = original_pil
+        # 1. Đọc ảnh
+        img = Image.open(image_path).convert("RGB")
+        img_np = np.array(img)
+
+        # 2. Chạy PaddleOCR (Local)
+        # Lưu ý: Model Local cơ bản không hỗ trợ tách bảng/công thức chuyên sâu như API
+        # nhưng ta vẫn chạy OCR để lấy text.
+        result = ocr.ocr(img_np, cls=True)
+
+        # 3. Xử lý kết quả để hiển thị
+        if not result or result[0] is None:
+            return "No text found.", "<p>No text detected</p>", "[]"
+
+        # Tạo ảnh visualization (Vẽ box)
+        annotated_img = universal_draw(img, result, FONT_PATH)
+        html_vis = pil_to_base64_html(annotated_img)
+
+        # Tạo Markdown Output
+        # Gom nhóm text lại thành đoạn văn
+        texts = [line[1][0] for line in result[0]]
         
-        # Kiểm tra xem Paddle có chỉnh sửa ảnh không (dựa vào key 'doc_preprocessor_res')
-        if isinstance(raw_result, list) and len(raw_result) > 0 and isinstance(raw_result[0], dict):
-            if 'doc_preprocessor_res' in raw_result[0]:
-                proc_res = raw_result[0]['doc_preprocessor_res']
-                # Nếu có ảnh đầu ra đã chỉnh sửa (output_img)
-                if 'output_img' in proc_res:
-                    print("Phát hiện ảnh đã qua xử lý hình học. Đang đồng bộ tọa độ...")
-                    numpy_img = proc_res['output_img']
-                    target_image_for_drawing = Image.fromarray(numpy_img)
-
-        # 3. Vẽ lên ảnh ĐÚNG (Target Image)
-        annotated_image = universal_draw(target_image_for_drawing, raw_result, FONT_PATH)
-
-        # 4. Xử lý Text
-        all_texts = deep_extract_text(raw_result)
-        final_texts = clean_text_result(all_texts)
-        text_output = "\n".join(final_texts) if final_texts else "Không tìm thấy văn bản."
-
-        # Debug Info
-        debug_str = str(raw_result)[:1000]
-        debug_info = f"Used Image Source: {'Preprocessed' if target_image_for_drawing != original_pil else 'Original'}\nData Preview:\n{debug_str}..."
-
-        return annotated_image, text_output, debug_info
+        if mode == "Formula":
+            md_text = "### Recognized Formula (Raw Text):\n\n" + " ".join(texts)
+            md_text += "\n\n*(Note: Local generic OCR model cannot convert to LaTeX math syntax)*"
+        elif mode == "Table":
+            md_text = "### Recognized Table Content:\n\n" + "\n".join(texts)
+            md_text += "\n\n*(Note: Local generic OCR model does not reconstruct HTML structure)*"
+        else: # Document / Generic
+            md_text = "### Document Content:\n\n" + "\n".join(texts)
+
+        # Raw Data (JSON string để debug)
+        raw_json = json.dumps(result[0], ensure_ascii=False, indent=2)
+
+        return md_text, html_vis, raw_json
 
     except Exception as e:
         import traceback
-        return image, f"Lỗi: {str(e)}", traceback.format_exc()
+        err = traceback.format_exc()
+        return f"Error: {str(e)}", f"<p style='color:red'>{str(e)}</p>", err
+
+# Wrapper cho các Tab khác nhau
+def run_doc_parsing(file, *args):
+    return local_inference(file, mode="Document")
 
-# --- GIAO DIỆN ---
-with gr.Blocks(title="PaddleOCR Perfect Overlay") as iface:
-    gr.Markdown("## PaddleOCR Chinese - High Precision Overlay")
+def run_element_recognition(file, prompt_label, *args):
+    # prompt_label: "Formula Recognition", "Table Recognition", etc.
+    mode = prompt_label.split()[0] # Lấy từ đầu tiên (Formula/Table...)
+    return local_inference(file, mode=mode)
+
+def run_spotting(file, *args):
+    # Spotting giả lập: Trả về bounding boxes của text dưới dạng JSON
+    if not file: return "", "{}"
     
-    with gr.Row():
-        with gr.Column():
-            input_img = gr.Image(type="pil", label="Input Image")
-            submit_btn = gr.Button("RUN OCR", variant="primary")
-        
-        with gr.Column():
-            with gr.Tabs():
-                with gr.TabItem("🖼️ Kết quả Khớp Tọa Độ"):
-                    output_img = gr.Image(type="pil", label="Overlay Result")
-                with gr.TabItem("📝 Văn bản"):
-                    output_txt = gr.Textbox(label="Text Content", lines=15)
-                with gr.TabItem("🐞 Debug"):
-                    output_debug = gr.Textbox(label="Debug Info", lines=15)
-
-    submit_btn.click(
-        fn=predict,
-        inputs=input_img,
-        outputs=[output_img, output_txt, output_debug]
-    )
+    img = Image.open(file).convert("RGB")
+    result = ocr.ocr(np.array(img), cls=True)
+    
+    if not result or result[0] is None:
+        return "<p>No objects found</p>", "[]"
+
+    annotated_img = universal_draw(img, result, FONT_PATH)
+    html_vis = pil_to_base64_html(annotated_img)
+    
+    # Format lại JSON cho giống spotting
+    spotting_res = []
+    for line in result[0]:
+        spotting_res.append({
+            "label": "text_block",
+            "text": line[1][0],
+            "confidence": line[1][1],
+            "box": line[0]
+        })
+    
+    return html_vis, json.dumps(spotting_res, ensure_ascii=False, indent=2)
+
+
+# --- PHẦN 2: GIAO DIỆN (UI TỪ FILE 2) ---
+custom_css = """
+body, .gradio-container { font-family: "Noto Sans SC", sans-serif; }
+.app-header { text-align: center; margin-bottom: 20px; }
+.prompt-grid { display: flex; flex-wrap: wrap; gap: 8px; margin-top: 6px; }
+.prompt-grid button { height: 40px !important; }
+.notice { background: #f0f9ff; padding: 10px; border-radius: 8px; border: 1px solid #bae6fd; font-size: 14px; margin-bottom: 10px;}
+"""
+
+with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
+    gr.HTML("""
+    <div class="app-header">
+        <h1>PaddleOCR Local - Pro Interface</h1>
+        <p>Giao diện nâng cao chạy trên Backend Local (CPU)</p>
+    </div>
+    <div class="notice">
+        <strong>Lưu ý:</strong> Đây là phiên bản chạy model Local. 
+        Các tính năng như <em>Formula to Latex</em>, <em>Table to HTML</em> hay <em>Layout Analysis</em> 
+        chỉ trả về văn bản thô (Raw OCR) do giới hạn của model cài đặt cục bộ.
+    </div>
+    """)
+
+    with gr.Tabs():
+        # ===================== Tab 1: Document Parsing =====================
+        with gr.Tab("Document Parsing"):
+            with gr.Row():
+                with gr.Column(scale=5):
+                    file_doc = gr.File(label="Upload Image", type="filepath", file_types=["image"])
+                    btn_parse = gr.Button("Parse Document", variant="primary")
+                    # Các tùy chọn checkbox (Dummy - vì local model config đơn giản)
+                    with gr.Row():
+                        gr.Checkbox(label="Chart parsing (N/A)", value=False, interactive=False)
+                        gr.Checkbox(label="Doc unwarping (N/A)", value=False, interactive=False)
+
+                with gr.Column(scale=7):
+                    with gr.Tabs():
+                        with gr.Tab("Markdown Preview"):
+                            md_preview_doc = gr.Markdown()
+                        with gr.Tab("Visualization"):
+                            vis_image_doc = gr.HTML()
+                        with gr.Tab("Raw Data"):
+                            raw_doc = gr.Code(language="json")
+
+            btn_parse.click(run_doc_parsing, inputs=[file_doc], outputs=[md_preview_doc, vis_image_doc, raw_doc])
+
+        # ===================== Tab 2: Element-level Recognition =====================
+        with gr.Tab("Element-level Recognition"):
+            with gr.Row():
+                with gr.Column(scale=5):
+                    file_vl = gr.File(label="Upload Image", type="filepath", file_types=["image"])
+                    gr.Markdown("_(Chế độ này tối ưu cho từng thành phần riêng lẻ)_")
+                    
+                    with gr.Row(elem_classes=["prompt-grid"]):
+                        btn_ocr = gr.Button("Text Recognition", variant="secondary")
+                        btn_formula = gr.Button("Formula Recognition", variant="secondary")
+                    with gr.Row(elem_classes=["prompt-grid"]):
+                        btn_table = gr.Button("Table Recognition", variant="secondary")
+                        btn_seal = gr.Button("Seal Recognition", variant="secondary")
+
+                with gr.Column(scale=7):
+                    with gr.Tabs():
+                        with gr.Tab("Result"):
+                            md_preview_vl = gr.Markdown()
+                        with gr.Tab("Visualization"):
+                             vis_image_vl = gr.HTML()
+                        with gr.Tab("Raw Output"):
+                            md_raw_vl = gr.Code(language="json")
+
+            # Gán sự kiện cho các nút
+            for btn, label in [(btn_ocr, "Text"), (btn_formula, "Formula"), (btn_table, "Table"), (btn_seal, "Seal")]:
+                btn.click(
+                    fn=run_element_recognition,
+                    inputs=[file_vl, gr.State(label)],
+                    outputs=[md_preview_vl, vis_image_vl, md_raw_vl]
+                )
+
+        # ===================== Tab 3: Spotting =====================
+        with gr.Tab("Spotting"):
+            with gr.Row():
+                with gr.Column(scale=5):
+                    file_spot = gr.File(label="Upload Image", type="filepath", file_types=["image"])
+                    btn_run_spot = gr.Button("Run Spotting", variant="primary")
+                    gr.Markdown("_(Phát hiện vị trí văn bản)_")
+
+                with gr.Column(scale=7):
+                    with gr.Tabs():
+                        with gr.Tab("Visualization"):
+                            vis_image_spot = gr.HTML()
+                        with gr.Tab("JSON Result"):
+                            json_spot = gr.Code(language="json")
+
+            btn_run_spot.click(run_spotting, inputs=[file_spot], outputs=[vis_image_spot, json_spot])
 
 if __name__ == "__main__":
-    iface.launch(server_name="0.0.0.0", server_port=7860)
+    demo.queue().launch(server_name="0.0.0.0", server_port=7860)