Update app.py

app.py

@@ -1,83 +1,208 @@
-import pytesseract
+import gradio as gr
+import fitz  # PyMuPDF
 from PIL import Image
-import re
+import numpy as np
+import cv2
+import pytesseract
+import base64
+import os
 import unicodedata
 
-# Configuración avanzada de Tesseract para OCR en español
-OCR_CONFIG = r'--oem 3 --psm 6'
-
-# Corrección básica de errores comunes de OCR
-def clean_text(text):
-    # Normaliza caracteres Unicode (acentos, símbolos)
-    text = unicodedata.normalize('NFKC', text)
-
-    # Diccionario de correcciones frecuentes
-    replacements = {
-        "martphone": "smartphone",
-        "desinstatar": "desinstalar",
-        "Desconectas": "desconectadas",
-        "cuestiónn": "cuestión",
-        "el coche sólo cargando": "el coche solo carga cuando",
-        "Modo online": "Modo en línea",
-        "Modo Online": "Modo en línea",
-        "Modo sin conexión": "modo sin conexión",
-        "desconectas": "desconectadas",
-        "cuestión no resolverá": "cuestión no solucionará",
-        "El símbolo del globo": "\nEl símbolo del globo",
-    }
-
-    for wrong, correct in replacements.items():
-        text = text.replace(wrong, correct)
+# NUEVO: Traducción
+from transformers import pipeline
 
-    return text
+# Inicializa el pipeline de traducción EN->ES una sola vez
+translator = pipeline("translation_en_to_es", model="Helsinki-NLP/opus-mt-en-es")
+
+# ---------- OCR y limpieza de texto ----------
 
-# Formatea el texto limpio en Markdown estructurado
-def format_text_to_markdown(text):
-    text = clean_text(text)
+def clean_ocr_text(text):
+    text = unicodedata.normalize("NFC", text)
     lines = text.splitlines()
-    markdown = []
+    cleaned_lines = [line.strip() for line in lines if line.strip()]
+    return "\n".join(cleaned_lines)
 
-    for line in lines:
-        line = line.strip()
-        if not line:
-            continue
+def translate_text(text):
+    """
+    Traduce texto del inglés al español si está en inglés (siempre lo traduce para simplificar)
+    """
+    # Para hacerlo robusto podrías agregar detección de idioma (langdetect),
+    # pero para este ejemplo traducimos siempre
+    if len(text.strip()) < 5:
+        return text
+    chunks = [text[i:i+500] for i in range(0, len(text), 500)]
+    translated = []
+    for chunk in chunks:
+        result = translator(chunk)
+        translated.append(result[0]["translation_text"])
+    return "\n".join(translated)
 
-        # Encabezados comunes detectados por OCR
-        if re.search(r"^posible causa", line, re.IGNORECASE):
-            markdown.append("### 🛑 Posible causa\n")
-            continue
-        elif re.search(r"^posible solución", line, re.IGNORECASE):
-            markdown.append("### ✅ Posible solución\n")
-            continue
-        elif re.search(r"^descripción del problema", line, re.IGNORECASE):
-            markdown.append("### 📝 Descripción del problema\n")
-            continue
+# ---------- Funciones de imagen ----------
 
-        # Listas con viñetas o números
-        elif re.match(r"^\d+\.", line):
-            markdown.append(f"- {line}")
-            continue
-        elif re.match(r"^[•*-]", line):
-            markdown.append(f"- {line[1:].strip()}")
+def text_area_ratio(image):
+    np_img = np.array(image.convert("L"))
+    _, thresh = cv2.threshold(np_img, 150, 255, cv2.THRESH_BINARY_INV)
+    contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    text_area = sum(w * h for x, y, w, h in [cv2.boundingRect(c) for c in contours if 8 < cv2.boundingRect(c)[3] < 40 and 5 < cv2.boundingRect(c)[2] < 100])
+    total_area = np_img.shape[0] * np_img.shape[1]
+    return text_area / total_area if total_area > 0 else 0
+
+def has_significant_text(image):
+    return text_area_ratio(image) > 0.25
+
+def is_primarily_text(image, ocr_threshold=30):
+    if has_significant_text(image):
+        ocr_result = pytesseract.image_to_string(image, lang="eng+spa")
+        return len(ocr_result.strip()) > ocr_threshold
+    return False
+
+def is_likely_photo(crop):
+    np_crop = np.array(crop)
+    gray = cv2.cvtColor(np_crop, cv2.COLOR_RGB2GRAY)
+    return np.std(gray) > 25 and len(np.unique(gray)) > 50
+
+def extract_visual_regions(image):
+    np_img = np.array(image.convert("RGB"))
+    gray = cv2.cvtColor(np_img, cv2.COLOR_RGB2GRAY)
+    _, binary = cv2.threshold(gray, 220, 255, cv2.THRESH_BINARY_INV)
+    closed = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, cv2.getStructuringElement(cv2.MORPH_RECT, (15, 15)))
+
+    num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(closed, connectivity=8)
+    results = []
+    for i in range(1, num_labels):
+        x, y, w, h, area = stats[i]
+        if area > 2000 and 0.3 < (w / float(h)) < 3.5:
+            bbox = (x, y, x + w, y + h)
+            crop = image.crop(bbox)
+            if is_likely_photo(crop) and text_area_ratio(crop) < 0.25 and not is_primarily_text(crop):
+                results.append(crop)
+    return results
+
+# ---------- Extracción de texto + imágenes ----------
+
+def clean_bullet_line(text):
+    text = unicodedata.normalize("NFKC", text)
+    text = text.replace("e@", "-")
+    text = text.replace("@", "-")
+    text = text.replace("•", "-")
+    text = text.replace("*", "-")
+    text = text.replace("·", "-")
+    text = text.replace("–", "-")
+    text = " ".join(text.split())
+    return text
+
+def extract_text_markdown(doc, image_paths, page_index, seen_xrefs):
+    markdown_output = f"\n## Página {page_index + 1}\n\n"
+    image_counter = 1
+    elements = []
+    page = doc[0]
+    blocks = page.get_text("dict")["blocks"]
+
+    for b in blocks:
+        y = b["bbox"][1]
+        if b["type"] == 0:
+            for line in b["lines"]:
+                line_y = line["bbox"][1]
+                line_text = " ".join([span["text"] for span in line["spans"]]).strip()
+                line_text = clean_bullet_line(line_text)
+                max_font_size = max([span.get("size", 10) for span in line["spans"]])
+                if line_text:
+                    elements.append((line_y, line_text, max_font_size))
+
+    images_on_page = page.get_images(full=True)
+    for img_index, img in enumerate(images_on_page):
+        xref = img[0]
+        if xref in seen_xrefs:
             continue
+        seen_xrefs.add(xref)
+        try:
+            base_image = page.parent.extract_image(xref)
+            image_bytes = base_image["image"]
+            ext = base_image["ext"]
+            image_path = f"/tmp/imagen_p{page_index + 1}_{img_index + 1}.{ext}"
+            with open(image_path, "wb") as f:
+                f.write(image_bytes)
+            image_paths.append(image_path)
+            elements.append((float("inf") - img_index, f"\n\n![imagen_{image_counter}]({image_path})\n", 10))
+            image_counter += 1
+        except Exception as e:
+            elements.append((float("inf"), f"[Error imagen: {e}]", 10))
+
+    elements.sort(key=lambda x: x[0])
+    previous_y = None
+
+    for y, text, font_size in elements:
+        is_header = font_size >= 14
+        if previous_y is not None and abs(y - previous_y) > 10:
+            markdown_output += "\n"
+        translated = translate_text(text.strip())
+        markdown_output += f"\n### {translated}\n" if is_header else translated + "\n"
+        previous_y = y
+
+    markdown_output += "\n---\n\n"
+    return markdown_output.strip()
+
+# ---------- Función principal ----------
 
-        # Texto normal
-        markdown.append(line)
+def convert(pdf_file):
+    temp_pdf_path = pdf_file.name
+    doc = fitz.open(temp_pdf_path)
+    markdown_output = ""
+    image_paths = []
+    seen_xrefs = set()
 
-    return "\n\n".join(markdown)
+    for page_num in range(len(doc)):
+        page = doc[page_num]
+        text = page.get_text("text").strip()
 
-# Función principal: OCR → Limpieza → Markdown
-def ocr_to_markdown(image_path):
-    image = Image.open(image_path)
-    raw_text = pytesseract.image_to_string(image, lang='spa', config=OCR_CONFIG)
-    return format_text_to_markdown(raw_text)
+        if len(text) > 30:
+            # Texto nativo del PDF
+            extracted = extract_text_markdown([page], image_paths, page_num, seen_xrefs)
+            markdown_output += extracted + "\n"
+        else:
+            # Página "escaneada" -> OCR
+            markdown_output += f"\n## Página {page_num + 1}\n\n"
+            pix = page.get_pixmap(dpi=300)
+            img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
+            image_path = f"/tmp/ocr_page_{page_num + 1}.jpg"
+            img.save(image_path)
+            image_paths.append(image_path)
+            markdown_output += f"![imagen_pagina_{page_num + 1}]({image_path})\n"
 
-if __name__ == "__main__":
-    image_path = "pagina1_ocr.png"  # Cambia esto por tu nombre real
-    markdown_output = ocr_to_markdown(image_path)
+            try:
+                ocr_text = pytesseract.image_to_string(img, lang="eng+spa")
+            except pytesseract.TesseractError:
+                ocr_text = ""
+            ocr_text_clean = clean_ocr_text(ocr_text)
+            translated_ocr = translate_text(ocr_text_clean)
+            markdown_output += translated_ocr + "\n"
 
-    with open("resultado.md", "w", encoding="utf-8") as f:
-        f.write("## Página 1\n\n")
-        f.write(f"![imagen_pagina_1]({image_path})\n\n")
+            crops = extract_visual_regions(img)
+            for i, crop in enumerate(crops):
+                crop_path = f"/tmp/recorte_p{page_num + 1}_{i + 1}.jpg"
+                crop.save(crop_path)
+                image_paths.append(crop_path)
+                markdown_output += f"\n\n![imagen_detectada]({crop_path})\n"
+
+            markdown_output += "\n---\n\n"
+
+    markdown_path = "/tmp/resultado.md"
+    with open(markdown_path, "w", encoding="utf-8") as f:
         f.write(markdown_output)
 
+    return markdown_output.strip(), image_paths, markdown_path
+
+# ---------- Gradio Interface ----------
+
+with gr.Blocks() as demo:
+    with gr.Row():
+        pdf_input = gr.File(label="Upload your PDF", type="filepath", file_types=[".pdf"])
+        submit_btn = gr.Button("Process PDF")
+
+    markdown_output = gr.Textbox(label="Generated Markdown", lines=25, interactive=True)
+    gallery_output = gr.Gallery(label="Extracted and Detected Images", type="file")
+    download_md = gr.File(label="Download Markdown File")
+
+    submit_btn.click(fn=convert, inputs=[pdf_input], outputs=[markdown_output, gallery_output, download_md])
+
+demo.launch()