ToomBComicCleanerGradio.py

# app.py
import os
import shutil
import tempfile
import cv2
import numpy as np
import gradio as gr
from paddleocr import PaddleOCR

ocr = PaddleOCR(use_angle_cls=True, lang='en', det_model_dir='models/det', rec_model_dir='models/rec', cls_model_dir='models/cls')

def classify_background_color(avg_color, white_thresh=230, black_thresh=50, yellow_thresh=100):
    r, g, b = avg_color
    if r >= white_thresh and g >= white_thresh and b >= white_thresh:
        return (255, 255, 255)
    if r <= black_thresh and g <= black_thresh and b <= black_thresh:
        return (0, 0, 0)
    if r >= yellow_thresh and g >= yellow_thresh and b < yellow_thresh:
        return (255, 255, 0)
    return None

def sample_border_color(image, box, padding=2):
    h, w = image.shape[:2]
    x_min, y_min, x_max, y_max = box
    x_min = max(0, x_min - padding)
    x_max = min(w-1, x_max + padding)
    y_min = max(0, y_min - padding)
    y_max = min(h-1, y_max + padding)

    top = image[y_min:y_min+padding, x_min:x_max]
    bottom = image[y_max-padding:y_max, x_min:x_max]
    left = image[y_min:y_max, x_min:x_min+padding]
    right = image[y_min:y_max, x_max-padding:x_max]

    border_pixels = np.vstack((top.reshape(-1, 3), bottom.reshape(-1, 3),
                               left.reshape(-1, 3), right.reshape(-1, 3)))
    if border_pixels.size == 0:
        return (255, 255, 255)
    median_color = np.median(border_pixels, axis=0)
    return tuple(map(int, median_color))

def detect_text_boxes(image):
    results = ocr.ocr(image, cls=True)
    if not results or not results[0]:
        return []
    boxes = []
    for line in results[0]:
        box, (text, confidence) = line
        if text.strip():
            x_min = int(min(pt[0] for pt in box))
            x_max = int(max(pt[0] for pt in box))
            y_min = int(min(pt[1] for pt in box))
            y_max = int(max(pt[1] for pt in box))
            boxes.append(((x_min, y_min, x_max, y_max), text, confidence))
    return boxes

def remove_text_dynamic_fill(img_path, output_path):
    image = cv2.imread(img_path)
    if image is None:
        return

    if len(image.shape) == 2:
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
    elif image.shape[2] == 1:
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
    else:
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

    boxes = detect_text_boxes(image)

    for (bbox, text, confidence) in boxes:
        if confidence < 0.4 or not text.strip():
            continue

        x_min, y_min, x_max, y_max = bbox
        height = y_max - y_min

        if height <= 30:
            padding = 2
        elif height <= 60:
            padding = 4
        else:
            padding = 6

        x_min_p = max(0, x_min - padding)
        y_min_p = max(0, y_min - padding)
        x_max_p = min(image.shape[1]-1, x_max + padding)
        y_max_p = min(image.shape[0]-1, y_max + padding)

        sample_crop = image[y_min_p:y_max_p, x_min_p:x_max_p]
        avg_color = np.mean(sample_crop.reshape(-1, 3), axis=0)

        fill_color = classify_background_color(avg_color)
        if fill_color is None:
            fill_color = sample_border_color(image, (x_min, y_min, x_max, y_max))

        cv2.rectangle(image, (x_min_p, y_min_p), (x_max_p, y_max_p), fill_color, -1)

    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
    cv2.imwrite(output_path, image)

def process_folder(input_files):
    temp_output = tempfile.mkdtemp()
    for file in input_files:
        filename = os.path.basename(file.name)
        output_path = os.path.join(temp_output, filename)
        remove_text_dynamic_fill(file.name, output_path)

    zip_path = shutil.make_archive(temp_output, 'zip', temp_output)
    return zip_path

demo = gr.Interface(
    fn=process_folder,
    inputs=gr.File(file_types=[".jpg", ".jpeg", ".png"], file_count="multiple", label="Upload Comic Images"),
    outputs=gr.File(label="Download Cleaned Zip"),
    title="Comic Text Cleaner",
    description="Upload comic images and get a zip of cleaned versions (text removed). Uses PaddleOCR for detection."
)

demo.launch()