#!/usr/bin/env python3
"""hough_deskew.py — fine-skew angle detector for the archivdms OCR pipeline.

Sidecar script for internal/ocr/ocr.go's `hough` deskew method
(config.OCRConfig.DeskewMethod == "hough"), an ALTERNATIVE to the default
ImageMagick `-deskew` peak/valley text-line projection analysis
(deskewImage() in ocr.go). ImageMagick's approach needs surrounding
background/margin to find the page's background rows/columns and fails on
tightly-cropped phone photos of receipts (no margin context) — see
project_deskew_disable_for_photos_tested_negative and
project_deskew_border_trick_tested_negative in agent memory for two
previously-tried and rejected workarounds. This script separates ANGLE
DETECTION (via OpenCV, this file) from angle APPLICATION (plain `convert
-rotate <deg>` in ocr.go) per the recommendation that produced this rewrite.

Usage:
    python3 hough_deskew.py <image-path>

Behavior:
    - Reads the image with OpenCV, grayscale + Otsu threshold.
    - Finds the largest contour by area and takes cv2.minAreaRect() of it.
      This is deliberately NOT text-line-projection-based (that is exactly
      what ImageMagick already does and what fails on cropped photos) —
      minAreaRect degrades gracefully to "the boundary of whatever content is
      in frame" even when that content fills the whole image, which is
      normally the case for a tightly-cropped phone photo.
    - Falls back to cv2.HoughLinesP() long-line-angle voting if no usable
      contour is found (e.g. near-blank background, no single dominant
      shape) — takes the median angle of detected line segments within
      +/-45 degrees of horizontal.
    - Prints exactly one float (the skew angle in degrees, ImageMagick
      `-rotate` sign convention: positive = clockwise) to stdout and exits 0
      on success.
    - On any failure (bad path, unreadable image, no contours/lines found),
      prints nothing to stdout, writes a one-line reason to stderr, and
      exits non-zero. ocr.go's houghDeskewAngle treats this as "angle 0,
      keep going" — never a fatal OCR error.

Dependencies: opencv-python (or the Debian python3-opencv apt package, which
pulls in numpy as a transitive dependency) — no other third-party packages.
Deliberately not using the `deskew` PyPI package: it wraps a very similar
Radon/Hough approach but pulls in scikit-image, a much heavier dependency
tree, for no accuracy benefit found in testing.
"""
import sys

try:
    import cv2
    import numpy as np
except ImportError as exc:  # pragma: no cover - environment/dependency issue
    print(f"hough_deskew: missing dependency: {exc}", file=sys.stderr)
    sys.exit(2)


def _angle_from_min_area_rect(gray: "np.ndarray"):
    """Return a skew angle in degrees via Otsu threshold + largest contour's
    minAreaRect, or None if no usable contour was found."""
    _, thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
    contours, _ = cv2.findContours(thresh, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
    if not contours:
        return None

    largest = max(contours, key=cv2.contourArea)
    # Ignore contours covering too little of the frame — noise/artifacts, not
    # the document itself.
    img_area = gray.shape[0] * gray.shape[1]
    if cv2.contourArea(largest) < 0.05 * img_area:
        return None

    rect = cv2.minAreaRect(largest)
    angle = rect[2]  # OpenCV: angle in (-90, 0] for cv2.minAreaRect

    # Normalize to the smallest rotation that would make the rect's long side
    # horizontal (matches ImageMagick -deskew / -rotate's small-angle
    # convention rather than cv2's raw (-90, 0] range).
    w, h = rect[1]
    if w < h:
        angle = angle + 90
    if angle > 45:
        angle -= 90
    elif angle < -45:
        angle += 90
    return angle


def _angle_from_hough_lines(gray: "np.ndarray"):
    """Fallback: median angle of long line segments detected via
    HoughLinesP, restricted to +/-45 degrees of horizontal. Returns None if
    no usable lines were found."""
    edges = cv2.Canny(gray, 50, 150, apertureSize=3)
    lines = cv2.HoughLinesP(
        edges, 1, np.pi / 180, threshold=100, minLineLength=gray.shape[1] // 4, maxLineGap=20
    )
    if lines is None or len(lines) == 0:
        return None

    angles = []
    for line in lines:
        x1, y1, x2, y2 = line[0]
        dx, dy = x2 - x1, y2 - y1
        if dx == 0:
            continue
        angle = np.degrees(np.arctan2(dy, dx))
        if -45 <= angle <= 45:
            angles.append(angle)
    if not angles:
        return None
    return float(np.median(angles))


def main() -> int:
    if len(sys.argv) != 2:
        print("hough_deskew: usage: hough_deskew.py <image-path>", file=sys.stderr)
        return 2

    path = sys.argv[1]
    img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
    if img is None:
        print(f"hough_deskew: could not read image: {path}", file=sys.stderr)
        return 1

    angle = _angle_from_min_area_rect(img)
    if angle is None:
        angle = _angle_from_hough_lines(img)
    if angle is None:
        print("hough_deskew: no usable contour or line angle found", file=sys.stderr)
        return 1

    print(f"{angle:.4f}")
    return 0


if __name__ == "__main__":
    sys.exit(main())