detect.py

"""
Watermark auto-detection module.

Two strategies:
  1. Text detection via EasyOCR (optional install).
  2. Contrast-anomaly detection for semi-transparent logos/patterns (always available).
"""

import cv2
import numpy as np
from typing import List, Dict

try:
    import easyocr
    _reader_instance = None
    EASYOCR_AVAILABLE = True
except ImportError:
    EASYOCR_AVAILABLE = False


def _get_reader():
    global _reader_instance
    if _reader_instance is None:
        import easyocr
        _reader_instance = easyocr.Reader(["en"], gpu=False)
    return _reader_instance


def detect_watermarks(image_path: str) -> List[Dict]:
    """
    Detect watermarks in an image using available methods.

    Returns a list of region dicts:
        {x, y, w, h, confidence, type}
    All coordinates are in original image pixel space.
    """
    img = cv2.imread(image_path)
    if img is None:
        return []

    regions: List[Dict] = []

    if EASYOCR_AVAILABLE:
        regions.extend(_detect_text(img))

    regions.extend(_detect_transparent(img))
    return _merge_overlapping(regions)


# ---------------------------------------------------------------------------
# Text detection
# ---------------------------------------------------------------------------

def _detect_text(img: np.ndarray) -> List[Dict]:
    reader = _get_reader()
    h, w = img.shape[:2]
    results = reader.readtext(img, paragraph=False, min_size=10)

    regions = []
    for bbox, text, confidence in results:
        if confidence < 0.3:
            continue
        xs = [pt[0] for pt in bbox]
        ys = [pt[1] for pt in bbox]
        x = max(0, int(min(xs)) - 5)
        y = max(0, int(min(ys)) - 5)
        x2 = min(w, int(max(xs)) + 5)
        y2 = min(h, int(max(ys)) + 5)
        regions.append({
            "x": x, "y": y,
            "w": x2 - x, "h": y2 - y,
            "confidence": float(confidence),
            "type": "text",
            "label": text[:20],
        })
    return regions


# ---------------------------------------------------------------------------
# Semi-transparent / logo detection
# ---------------------------------------------------------------------------

def _detect_transparent(img: np.ndarray) -> List[Dict]:
    """
    Detect logo/pattern watermarks by finding structured residuals
    that deviate from the smoothed background.
    """
    h, w = img.shape[:2]
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY).astype(np.float32)

    # Background estimate via heavy blur
    bg = cv2.GaussianBlur(gray, (51, 51), 0)
    residual = np.abs(gray - bg).astype(np.uint8)

    _, thresh = cv2.threshold(residual, 15, 255, cv2.THRESH_BINARY)

    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
    closed = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel, iterations=3)

    n_labels, _, stats, _ = cv2.connectedComponentsWithStats(closed)

    min_area = h * w * 0.005
    max_area = h * w * 0.40

    regions = []
    for i in range(1, n_labels):
        area = stats[i, cv2.CC_STAT_AREA]
        if not (min_area <= area <= max_area):
            continue

        bx = stats[i, cv2.CC_STAT_LEFT]
        by = stats[i, cv2.CC_STAT_TOP]
        bw = stats[i, cv2.CC_STAT_WIDTH]
        bh = stats[i, cv2.CC_STAT_HEIGHT]

        aspect = bw / bh if bh > 0 else 0
        if aspect > 10 or aspect < 0.1:
            continue

        regions.append({
            "x": int(bx), "y": int(by),
            "w": int(bw), "h": int(bh),
            "confidence": 0.55,
            "type": "logo",
        })

    return regions


# ---------------------------------------------------------------------------
# Merge overlapping boxes
# ---------------------------------------------------------------------------

def _merge_overlapping(regions: List[Dict]) -> List[Dict]:
    if len(regions) <= 1:
        return regions

    boxes = [(r["x"], r["y"], r["x"] + r["w"], r["y"] + r["h"]) for r in regions]

    changed = True
    while changed:
        changed = False
        out: List = []
        used = [False] * len(boxes)

        for i in range(len(boxes)):
            if used[i]:
                continue
            x1, y1, x2, y2 = boxes[i]

            for j in range(i + 1, len(boxes)):
                if used[j]:
                    continue
                bx1, by1, bx2, by2 = boxes[j]
                if x1 < bx2 and x2 > bx1 and y1 < by2 and y2 > by1:
                    x1, y1 = min(x1, bx1), min(y1, by1)
                    x2, y2 = max(x2, bx2), max(y2, by2)
                    used[j] = True
                    changed = True

            out.append((x1, y1, x2, y2))
            used[i] = True

        boxes = out

    return [
        {"x": x1, "y": y1, "w": x2 - x1, "h": y2 - y1, "confidence": 0.7, "type": "merged"}
        for x1, y1, x2, y2 in boxes
    ]