app.py

from detect_bubbles import detect_bubbles
from process_bubble import extract_bubble_mask
from add_text import add_text
from translator import MangaTranslator
from manga_ocr import MangaOcr

from PIL import Image
import numpy as np
import cv2
import gradio as gr

MODEL = "model.pt"

# Modelos carregados uma única vez
manga_translator = MangaTranslator()
mocr = MangaOcr()


def predict(img, translation_method, font):
    # =========================
    # 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, original_image)

    page_mask = np.zeros(original_image.shape[:2], dtype=np.uint8)
    bubbles_data = []

    # =========================
    # 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_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({
            "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
        )

        render_image[y1:y2, x1:x2] = add_text(
            render_image[y1:y2, x1:x2],
            translated,
            font,
            contour
        )

    return Image.fromarray(render_image)


# =========================
# Interface Gradio
# =========================

demo = gr.Interface(
    fn=predict,
    inputs=[
        "image",
        gr.Dropdown(
            [
                ("Google", "google"),
                ("HuggingFace", "hf"),
                ("Bing", "bing"),
                ("Sogou", "sogou")
            ],
            value="google",
            label="Translation Method"
        ),
        gr.Dropdown(
            [
                ("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",
    description="Translate manga speech bubbles safely and correctly"
)

demo.launch(debug=False, share=False)