File size: 21,631 Bytes
5196d55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
#!/usr/bin/env python3
"""
Popescu–Farid CFA Interpolation Forensics (End-to-End)

Implements the algorithm from:

  A. C. Popescu and H. Farid,
  "Exposing Digital Forgeries in Color Filter Array Interpolated Images,"
  IEEE Trans. Signal Processing, 2005.

Features:
  - EM estimation of the linear correlation model for each color channel
    (single-channel Gaussian vs uniform mixture).
  - Posterior probability map per channel.
  - Synthetic CFA maps s_r, s_g, s_b for a Bayer pattern.
  - Fourier-domain similarity M(p_c, s_c) per channel.
  - Sliding-window analysis with 50% overlap.
  - Threshold calibration on a set of negative images to get ~0% FPs.
  - Multi-channel fusion (default): window authentic if ANY channel is CFA;
    optional green-only mode: window authentic if GREEN channel is CFA.

Usage:

  # 1) Calibrate thresholds on negative (non-CFA / tampered) images
  python pf_cfa_detector.py calibrate \
      --neg-dir path/to/negative_images \
      --output thresholds.json

  # 2) Detect on a single image (all channels, PF-style fusion)
  python pf_cfa_detector.py detect \
      --image path/to/test_image.png \
      --thresholds thresholds.json

  # 3) Detect on a directory, green-only mode
  python pf_cfa_detector.py detect \
      --image-dir path/to/test_images \
      --thresholds thresholds.json \
      --green-only
"""

import argparse
import json
import math
import os
from typing import Dict, List, Sequence, Tuple

import numpy as np
from numpy.fft import fft2
from PIL import Image


# -------------------------------------------------------------------------
# Parameters and utilities
# -------------------------------------------------------------------------

IMG_EXTS = (".jpg", ".jpeg", ".png", ".bmp", ".tif", ".tiff")


def list_image_files(directory: str) -> List[str]:
    """Return all image files in a directory (non-recursive) with known extensions."""
    paths = []
    for name in os.listdir(directory):
        if name.lower().endswith(IMG_EXTS):
            paths.append(os.path.join(directory, name))
    paths.sort()
    return paths


def load_rgb_image(path: str) -> np.ndarray:
    """
    Load an image from disk and return as float64 RGB array in [0,1].

    Parameters
    ----------
    path : str
        Path to image file.

    Returns
    -------
    img : np.ndarray, shape (H, W, 3), dtype float64
        RGB image, intensities in [0, 1].
    """
    im = Image.open(path).convert("RGB")
    arr = np.asarray(im, dtype=np.float64)
    if arr.ndim != 3 or arr.shape[2] != 3:
        raise ValueError(f"Expected an RGB image at {path}")
    return arr / 255.0


# -------------------------------------------------------------------------
# EM algorithm for the linear CFA correlation model (single channel)
# -------------------------------------------------------------------------

def em_probability_map(
    f: np.ndarray,
    N: int = 1,
    sigma0: float = 0.0075,
    p0: float = 1.0 / 256.0,
    max_iter: int = 50,
    tol: float = 1e-5,
    seed: int = 0,
) -> Tuple[np.ndarray, np.ndarray]:
    """
    Estimate the posterior probability map w(x,y) that each sample belongs to
    the linearly correlated model M1 via EM (Gaussian vs uniform mixture).

    Model for a single channel f(x,y):
        f(x,y) = sum_{u,v} alpha_{u,v} f(x+u, y+v) + n(x,y)
        where n(x,y) ~ N(0, sigma^2), alpha_{0,0} = 0.
    """

    if f.ndim != 2:
        raise ValueError("em_probability_map expects a single 2D channel.")

    f = f.astype(np.float64, copy=False)
    H, W = f.shape
    if H <= 2 * N or W <= 2 * N:
        raise ValueError("Image too small for N = %d neighborhood." % N)

    # For N=1, there are (2N+1)^2 - 1 = 8 neighbors.
    OFFSETS: List[Tuple[int, int]] = []
    for dy in range(-N, N + 1):
        for dx in range(-N, N + 1):
            if dy == 0 and dx == 0:
                continue
            OFFSETS.append((dy, dx))
    K = len(OFFSETS)

    # Build design matrix X and observation vector y.
    num_pixels = (H - 2 * N) * (W - 2 * N)
    X = np.empty((num_pixels, K), dtype=np.float64)
    y = np.empty((num_pixels, 1), dtype=np.float64)

    idx = 0
    for yy in range(N, H - N):
        for xx in range(N, W - N):
            y[idx, 0] = f[yy, xx]
            for k, (dy, dx) in enumerate(OFFSETS):
                X[idx, k] = f[yy + dy, xx + dx]
            idx += 1

    rng = np.random.default_rng(seed)
    alpha = rng.normal(scale=0.01, size=(K, 1))

    # ---- NEW: enforce lower bound on sigma to avoid degeneracy ----
    EPS_SIGMA = 1e-6
    sigma = float(max(sigma0, EPS_SIGMA))

    for _ in range(max_iter):
        # E-step: residuals and posterior weights
        r = y - X @ alpha  # (num_pixels,1)

        # Guard sigma here too
        if sigma < EPS_SIGMA:
            sigma = EPS_SIGMA

        coef = 1.0 / (sigma * math.sqrt(2.0 * math.pi))
        P = coef * np.exp(-0.5 * (r / sigma) ** 2)  # Gaussian likelihood for M1

        # Posterior w = P / (P + p0)
        w = P / (P + p0)

        # M-step: weighted LS for alpha
        WX = X * w
        A = X.T @ WX
        b = X.T @ (w * y)

        try:
            alpha_new = np.linalg.solve(A, b)
        except np.linalg.LinAlgError:
            alpha_new, *_ = np.linalg.lstsq(A, b, rcond=None)

        # Update sigma^2 = sum w_i r_i^2 / sum w_i
        r = y - X @ alpha_new
        num = float(np.sum(w * (r ** 2)))
        den = float(np.sum(w))

        if den > 0.0:
            sigma_new = math.sqrt(num / den)
        else:
            sigma_new = sigma

        # Clamp again to avoid zero
        if sigma_new < EPS_SIGMA:
            sigma_new = EPS_SIGMA

        # Convergence check
        diff = np.linalg.norm(alpha_new - alpha)
        norm = np.linalg.norm(alpha)
        if norm > 0 and diff < tol * norm:
            alpha = alpha_new
            sigma = sigma_new
            break

        alpha = alpha_new
        sigma = sigma_new

    # Final posterior with converged alpha
    if sigma < EPS_SIGMA:
        sigma = EPS_SIGMA

    r = y - X @ alpha
    coef = 1.0 / (sigma * math.sqrt(2.0 * math.pi))
    P = coef * np.exp(-0.5 * (r / sigma) ** 2)
    w = P / (P + p0)

    prob_map = np.zeros_like(f)
    idx = 0
    for yy in range(N, H - N):
        for xx in range(N, W - N):
            prob_map[yy, xx] = w[idx, 0]
            idx += 1

    return prob_map, alpha.ravel()



# -------------------------------------------------------------------------
# CFA synthetic maps and Fourier-domain similarity
# -------------------------------------------------------------------------

def synthetic_cfa_map(
    shape: Tuple[int, int],
    channel: str,
    pattern: str = "RGGB",
) -> np.ndarray:
    """
    Build the synthetic binary map s_c(x,y) for a Bayer pattern:

        s_c(x,y) = 0 if CFA at (x,y) is color c
                 = 1 otherwise

    Bayer patterns are specified as a 4-character string:
        'RGGB', 'BGGR', 'GRBG', 'GBRG', etc.

    pattern[0] -> (row%2==0, col%2==0)
    pattern[1] -> (row%2==0, col%2==1)
    pattern[2] -> (row%2==1, col%2==0)
    pattern[3] -> (row%2==1, col%2==1)
    """
    H, W = shape
    if len(pattern) != 4:
        raise ValueError("Bayer pattern string must have length 4, e.g. 'RGGB'.")

    channel = channel.upper()
    if channel not in ("R", "G", "B"):
        raise ValueError("channel must be one of 'R', 'G', 'B'.")

    tile = np.array(list(pattern), dtype="<U1").reshape(2, 2)
    s = np.ones((H, W), dtype=np.float64)

    for y in range(H):
        for x in range(W):
            cy = y % 2
            cx = x % 2
            if tile[cy, cx] == channel:
                s[y, x] = 0.0

    return s


def similarity_measure(prob_map: np.ndarray, synthetic_map: np.ndarray) -> float:
    """
    Phase-insensitive similarity between a probability map and its CFA
    synthetic map:

        M(p,s) = sum |F(p)| * |F(s)|
    """
    if prob_map.shape != synthetic_map.shape:
        raise ValueError("prob_map and synthetic_map must have the same shape.")

    Fp = fft2(prob_map)
    Fs = fft2(synthetic_map)
    return float(np.sum(np.abs(Fp) * np.abs(Fs)))


# -------------------------------------------------------------------------
# Sliding-window analysis
# -------------------------------------------------------------------------

def sliding_window_indices(H: int, W: int, window: int) -> List[Tuple[int, int]]:
    """
    Generate (y,x) indices for sliding windows with 50% overlap.
    stride = window // 2 along each axis.
    """
    if window > H or window > W:
        return [(0, 0)]

    stride = max(1, window // 2)
    indices = []
    y = 0
    while y + window <= H:
        x = 0
        while x + window <= W:
            indices.append((y, x))
            x += stride
        y += stride
    return indices


def analyze_window(
    window: np.ndarray,
    pattern: str = "RGGB",
    em_kwargs: Dict = None,
) -> Dict[str, Dict[str, np.ndarray]]:
    """
    Run EM + CFA similarity on a single RGB window.

    Returns per-channel:
        'prob_map', 'synthetic', 'M', 'alpha'
    """
    if window.ndim != 3 or window.shape[2] != 3:
        raise ValueError("Expected RGB window of shape (H,W,3).")

    if em_kwargs is None:
        em_kwargs = {}

    H, W, _ = window.shape
    channels = {
        "R": window[:, :, 0],
        "G": window[:, :, 1],
        "B": window[:, :, 2],
    }

    result: Dict[str, Dict[str, np.ndarray]] = {}
    for cname, ch in channels.items():
        prob_map, alpha = em_probability_map(ch, **em_kwargs)
        syn = synthetic_cfa_map((H, W), cname, pattern=pattern)
        M = similarity_measure(prob_map, syn)
        result[cname] = {
            "prob_map": prob_map,
            "synthetic": syn,
            "M": np.array(M, dtype=np.float64),
            "alpha": alpha,
        }

    return result


def analyze_image_windows(
    img: np.ndarray,
    window: int = 256,
    pattern: str = "RGGB",
    em_kwargs: Dict = None,
) -> List[Dict]:
    """
    Apply CFA EM analysis to all sliding windows of an image.

    If image is smaller than `window`, the entire image is treated as one window.
    """
    H, W, C = img.shape
    if C != 3:
        raise ValueError("Expected RGB image with 3 channels.")

    if em_kwargs is None:
        em_kwargs = {}

    if H < window or W < window:
        windows = [(0, 0)]
        w_h, w_w = H, W
    else:
        windows = sliding_window_indices(H, W, window)
        w_h = w_w = window

    results = []
    for (yy, xx) in windows:
        sub = img[yy : yy + w_h, xx : xx + w_w, :]
        res = analyze_window(sub, pattern=pattern, em_kwargs=em_kwargs)
        entry = {"y": yy, "x": xx, "h": sub.shape[0], "w": sub.shape[1]}
        entry.update(res)
        results.append(entry)

    return results


# -------------------------------------------------------------------------
# Threshold calibration and classification
# -------------------------------------------------------------------------

def calibrate_thresholds(
    negative_image_paths: Sequence[str],
    window: int = 256,
    pattern: str = "RGGB",
    em_kwargs: Dict = None,
) -> Dict[str, float]:
    """
    Estimate per-channel thresholds T_R, T_G, T_B to obtain ~0% false positives
    on a negative set (non-CFA / tampered images).

    Threshold per channel is defined as the maximum M value observed for that
    channel over all windows of all negative images.
    """
    if em_kwargs is None:
        em_kwargs = {}

    Ms = {"R": [], "G": [], "B": []}

    for path in negative_image_paths:
        img = load_rgb_image(path)
        window_results = analyze_image_windows(
            img,
            window=window,
            pattern=pattern,
            em_kwargs=em_kwargs,
        )
        for r in window_results:
            Ms["R"].append(float(r["R"]["M"]))
            Ms["G"].append(float(r["G"]["M"]))
            Ms["B"].append(float(r["B"]["M"]))

    thresholds: Dict[str, float] = {}
    for c in ("R", "G", "B"):
        values = Ms[c]
        if not values:
            raise RuntimeError(f"No M values collected for channel {c}.")
        thresholds[c] = max(values)

    return thresholds


def classify_windows(
    window_results: List[Dict],
    thresholds: Dict[str, float],
    green_only: bool = False,
) -> List[Dict]:
    """
    Add CFA/tampered labels to each window based on per-channel thresholds.

    If green_only=False (default, PF-style):
        channel is CFA-interpolated  <=>  M_c > T_c
        window authentic             <=>  any channel is CFA-interpolated

    If green_only=True:
        channel flags are still computed, but
        window authentic             <=>  GREEN channel is CFA-interpolated.
    """
    classified = []
    for r in window_results:
        M_R = float(r["R"]["M"])
        M_G = float(r["G"]["M"])
        M_B = float(r["B"]["M"])

        chan_cfa = {
            "R": M_R > thresholds["R"],
            "G": M_G > thresholds["G"],
            "B": M_B > thresholds["B"],
        }

        if green_only:
            authentic = chan_cfa["G"]
        else:
            authentic = chan_cfa["R"] or chan_cfa["G"] or chan_cfa["B"]

        out = dict(r)
        out["channel_cfa"] = chan_cfa
        out["authentic"] = authentic
        classified.append(out)

    return classified


def classify_image(
    img_path: str,
    thresholds: Dict[str, float],
    window: int = 256,
    pattern: str = "RGGB",
    em_kwargs: Dict = None,
    green_only: bool = False,
) -> Dict:
    """
    Run full sliding-window Popescu–Farid-style detector on a single image.

    Returns:
        {
          "image_path": str,
          "windows": [...],
          "image_authentic": bool,
        }
    """
    img = load_rgb_image(img_path)
    window_results = analyze_image_windows(
        img,
        window=window,
        pattern=pattern,
        em_kwargs=em_kwargs,
    )
    classified = classify_windows(window_results, thresholds, green_only=green_only)

    image_authentic = any(w["authentic"] for w in classified)

    return {
        "image_path": img_path,
        "windows": classified,
        "image_authentic": image_authentic,
    }


# -------------------------------------------------------------------------
# CLI plumbing
# -------------------------------------------------------------------------

def add_em_args(parser: argparse.ArgumentParser) -> None:
    """Add EM-related arguments to a subparser."""
    parser.add_argument(
        "--N",
        type=int,
        default=1,
        help="Neighborhood radius N for EM (default: 1).",
    )
    parser.add_argument(
        "--sigma0",
        type=float,
        default=0.0075,
        help="Initial sigma_0 for EM (default: 0.0075).",
    )
    parser.add_argument(
        "--p0",
        type=float,
        default=1.0 / 256.0,
        help=(
            "Outlier likelihood p0. Default 1/256 (PF-style for 8-bit data). "
            "For a uniform on [0,1], consider --p0 1.0."
        ),
    )
    parser.add_argument(
        "--max-iter",
        type=int,
        default=50,
        help="Maximum EM iterations (default: 50).",
    )
    parser.add_argument(
        "--tol",
        type=float,
        default=1e-5,
        help="Relative convergence tolerance on alpha (default: 1e-5).",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=0,
        help="Random seed for EM initialization (default: 0).",
    )


def em_args_to_kwargs(args: argparse.Namespace) -> Dict:
    """Convert parsed EM args into kwargs dict for em_probability_map."""
    return dict(
        N=args.N,
        sigma0=args.sigma0,
        p0=args.p0,
        max_iter=args.max_iter,
        tol=args.tol,
        seed=args.seed,
    )


def main():
    parser = argparse.ArgumentParser(
        description="Popescu–Farid CFA interpolation detector (EM + spectral similarity)."
    )
    subparsers = parser.add_subparsers(dest="command", required=True)

    # Calibrate subcommand
    cal_parser = subparsers.add_parser(
        "calibrate",
        help="Calibrate per-channel thresholds on negative (non-CFA) images.",
    )
    cal_parser.add_argument(
        "--neg-dir",
        required=True,
        help="Directory with negative (non-CFA / tampered) images for calibration.",
    )
    cal_parser.add_argument(
        "--window",
        type=int,
        default=256,
        help="Window size (square) for analysis (default: 256).",
    )
    cal_parser.add_argument(
        "--pattern",
        type=str,
        default="RGGB",
        help="Bayer pattern string (default: RGGB).",
    )
    cal_parser.add_argument(
        "--output",
        type=str,
        default=None,
        help="Path to JSON file to save thresholds (optional).",
    )
    add_em_args(cal_parser)

    # Detect subcommand
    det_parser = subparsers.add_parser(
        "detect",
        help="Run detector on images using pre-calibrated thresholds.",
    )
    det_parser.add_argument(
        "--image",
        type=str,
        default=None,
        help="Path to a single image.",
    )
    det_parser.add_argument(
        "--image-dir",
        type=str,
        default=None,
        help="Directory with images to process.",
    )
    det_parser.add_argument(
        "--thresholds",
        type=str,
        required=True,
        help="Path to JSON file with thresholds from 'calibrate'.",
    )
    det_parser.add_argument(
        "--window",
        type=int,
        default=256,
        help="Window size (square) for analysis (default: 256).",
    )
    det_parser.add_argument(
        "--pattern",
        type=str,
        default="RGGB",
        help="Bayer pattern string (default: RGGB).",
    )
    det_parser.add_argument(
        "--green-only",
        action="store_true",
        help="Use only GREEN channel for window/image decisions.",
    )
    add_em_args(det_parser)

    args = parser.parse_args()

    if args.command == "calibrate":
        neg_files = list_image_files(args.neg_dir)
        if not neg_files:
            raise SystemExit(f"No images found in negative directory: {args.neg_dir}")

        em_kwargs = em_args_to_kwargs(args)
        thresholds = calibrate_thresholds(
            negative_image_paths=neg_files,
            window=args.window,
            pattern=args.pattern,
            em_kwargs=em_kwargs,
        )

        # Print thresholds
        print("# Calibrated thresholds (0%% FPs on provided negatives):")
        for c in ("R", "G", "B"):
            print(f"T_{c} = {thresholds[c]:.6e}")

        # Optionally save to JSON
        if args.output is not None:
            out_obj = {
                "thresholds": thresholds,
                "pattern": args.pattern,
                "window": args.window,
                "em_params": em_kwargs,
            }
            with open(args.output, "w", encoding="utf-8") as f:
                json.dump(out_obj, f, indent=2)
            print(f"# Saved thresholds to {args.output}")

    elif args.command == "detect":
        # Load thresholds JSON
        with open(args.thresholds, "r", encoding="utf-8") as f:
            data = json.load(f)

        if "thresholds" in data:
            thresholds = data["thresholds"]
        else:
            thresholds = data  # assume raw

        # Sanity check
        for c in ("R", "G", "B"):
            if c not in thresholds:
                raise SystemExit(f"Thresholds JSON missing channel '{c}'.")

        # Decide images to process
        images: List[str] = []
        if args.image is not None and args.image_dir is not None:
            raise SystemExit("Specify either --image or --image-dir, not both.")
        elif args.image is not None:
            images = [args.image]
        elif args.image_dir is not None:
            images = list_image_files(args.image_dir)
            if not images:
                raise SystemExit(
                    f"No images found in directory: {args.image_dir}"
                )
        else:
            raise SystemExit("You must specify either --image or --image-dir.")

        em_kwargs = em_args_to_kwargs(args)
        green_only = bool(args.green_only)

        for img_path in images:
            result = classify_image(
                img_path,
                thresholds=thresholds,
                window=args.window,
                pattern=args.pattern,
                em_kwargs=em_kwargs,
                green_only=green_only,
            )

            label = "AUTHENTIC" if result["image_authentic"] else "TAMPERED"
            mode = "GREEN_ONLY" if green_only else "ALL_CHANNELS"
            print(f"\n# Image: {result['image_path']}")
            print(f"# Overall label: {label} (mode={mode})")
            print("# y x h w  M_R  M_G  M_B  CFA_R CFA_G CFA_B  WINDOW_LABEL")

            for w in result["windows"]:
                y = w["y"]
                x = w["x"]
                h = w["h"]
                wd = w["w"]
                M_R = float(w["R"]["M"])
                M_G = float(w["G"]["M"])
                M_B = float(w["B"]["M"])
                cfaR = w["channel_cfa"]["R"]
                cfaG = w["channel_cfa"]["G"]
                cfaB = w["channel_cfa"]["B"]
                wlabel = "AUTH" if w["authentic"] else "TAMP"

                print(
                    f"{y:5d} {x:5d} {h:4d} {wd:4d}  "
                    f"{M_R:.6e} {M_G:.6e} {M_B:.6e}  "
                    f"{int(cfaR)} {int(cfaG)} {int(cfaB)}  {wlabel}"
                )


if __name__ == "__main__":
    main()