Update app.py

app.py

@@ -11,69 +11,107 @@ import gradio as gr
 
 MODEL = "model.pt"
 
+# Modelos carregados uma única vez
 manga_translator = MangaTranslator()
 mocr = MangaOcr()
 
 
 def predict(img, translation_method, font):
-    image = np.array(img)
+    # =========================
+    # Imagens separadas por função
+    # =========================
+    original_image = np.array(img).copy()   # usada SOMENTE para OCR
+    render_image = np.array(img).copy()     # usada para limpeza + escrita
 
-    results = detect_bubbles(MODEL, image)
+    results = detect_bubbles(MODEL, original_image)
 
-    page_mask = np.zeros(image.shape[:2], dtype=np.uint8)
+    page_mask = np.zeros(original_image.shape[:2], dtype=np.uint8)
     bubbles_data = []
 
-    # 1️⃣ Extrai TODAS as máscaras primeiro
+    # =========================
+    # 1) OCR + extração de máscara (SEM modificar imagem)
+    # =========================
     for x1, y1, x2, y2, score, class_id in results:
         x1, y1, x2, y2 = map(int, (x1, y1, x2, y2))
-        crop = image[y1:y2, x1:x2]
 
-        mask, contour = extract_bubble_mask(crop)
+        crop_original = original_image[y1:y2, x1:x2]
+
+        # OCR SEMPRE na imagem original
+        pil_crop = Image.fromarray(crop_original.astype(np.uint8))
+        text = mocr(pil_crop)
+
+        mask, contour = extract_bubble_mask(crop_original)
         if mask is None:
             continue
 
         page_mask[y1:y2, x1:x2][mask == 255] = 255
-        bubbles_data.append((x1, y1, x2, y2, crop, contour))
 
-    # 2️⃣ Limpa a página UMA VEZ
-    image[page_mask == 255] = (255, 255, 255)
-
-    # 3️⃣ OCR + Tradução + Escrita
-    for x1, y1, x2, y2, crop, contour in bubbles_data:
-        pil_crop = Image.fromarray(crop.astype(np.uint8))
-        text = mocr(pil_crop)
+        bubbles_data.append({
+            "box": (x1, y1, x2, y2),
+            "text": text,
+            "contour": contour
+        })
+
+    # =========================
+    # 2) Limpeza global da página (UMA VEZ)
+    # =========================
+    render_image[page_mask == 255] = (255, 255, 255)
+
+    # =========================
+    # 3) Tradução + escrita
+    # =========================
+    for bubble in bubbles_data:
+        x1, y1, x2, y2 = bubble["box"]
+        text = bubble["text"]
+        contour = bubble["contour"]
 
         translated = manga_translator.translate(
-            text, method=translation_method
+            text,
+            method=translation_method
         )
 
-        image[y1:y2, x1:x2] = add_text(
-            image[y1:y2, x1:x2],
+        render_image[y1:y2, x1:x2] = add_text(
+            render_image[y1:y2, x1:x2],
             translated,
             font,
             contour
         )
 
-    return Image.fromarray(image)
+    return Image.fromarray(render_image)
+
 
+# =========================
+# Interface Gradio
+# =========================
 
 demo = gr.Interface(
     fn=predict,
     inputs=[
         "image",
         gr.Dropdown(
-            [("Google", "google"), ("HF", "hf")],
+            [
+                ("Google", "google"),
+                ("HuggingFace", "hf"),
+                ("Bing", "bing"),
+                ("Sogou", "sogou")
+            ],
             value="google",
-            label="Translation"
+            label="Translation Method"
         ),
         gr.Dropdown(
-            [("WildWords", "fonts/WildWordsRoman.ttf")],
+            [
+                ("Wild Words Roman", "fonts/WildWordsRoman.ttf"),
+                ("Anime Ace Italic", "fonts/animeace_i.ttf"),
+                ("Mangati", "fonts/mangati.ttf"),
+                ("Komika Axis", "fonts/KOMIKAX_.ttf")
+            ],
             value="fonts/WildWordsRoman.ttf",
             label="Font"
         )
     ],
     outputs=gr.Image(),
-    title="Manga Translator"
+    title="Manga Translator",
+    description="Translate manga speech bubbles safely and correctly"
 )
 
-demo.launch()
+demo.launch(debug=False, share=False)