File size: 7,493 Bytes
128f42e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
utils.py

Helper functions for image postprocessing, including EXIF removal, noise addition,
color correction, and Fourier spectrum matching.
"""

from PIL import Image, ImageOps
import numpy as np
try:
    import cv2
    _HAS_CV2 = True
except Exception:
    cv2 = None
    _HAS_CV2 = False
from scipy.ndimage import gaussian_filter1d

def remove_exif_pil(img: Image.Image) -> Image.Image:
    data = img.tobytes()
    new = Image.frombytes(img.mode, img.size, data)
    return new

def add_gaussian_noise(img_arr: np.ndarray, std_frac=0.02, seed=None) -> np.ndarray:
    if seed is not None:
        np.random.seed(seed)
    std = std_frac * 255.0
    noise = np.random.normal(loc=0.0, scale=std, size=img_arr.shape)
    out = img_arr.astype(np.float32) + noise
    out = np.clip(out, 0, 255).astype(np.uint8)
    return out

def clahe_color_correction(img_arr: np.ndarray, clip_limit=2.0, tile_grid_size=(8,8)) -> np.ndarray:
    if _HAS_CV2:
        lab = cv2.cvtColor(img_arr, cv2.COLOR_RGB2LAB)
        l, a, b = cv2.split(lab)
        clahe = cv2.createCLAHE(clipLimit=clip_limit, tileGridSize=tile_grid_size)
        l2 = clahe.apply(l)
        lab2 = cv2.merge((l2, a, b))
        out = cv2.cvtColor(lab2, cv2.COLOR_LAB2RGB)
        return out
    else:
        pil = Image.fromarray(img_arr)
        channels = pil.split()
        new_ch = []
        for ch in channels:
            eq = ImageOps.equalize(ch)
            new_ch.append(eq)
        merged = Image.merge('RGB', new_ch)
        return np.array(merged)

def randomized_perturbation(img_arr: np.ndarray, magnitude_frac=0.008, seed=None) -> np.ndarray:
    if seed is not None:
        np.random.seed(seed)
    mag = magnitude_frac * 255.0
    perturb = np.random.uniform(low=-mag, high=mag, size=img_arr.shape)
    out = img_arr.astype(np.float32) + perturb
    out = np.clip(out, 0, 255).astype(np.uint8)
    return out

def radial_profile(mag: np.ndarray, center=None, nbins=None):
    h, w = mag.shape
    if center is None:
        cy, cx = h // 2, w // 2
    else:
        cy, cx = center

    if nbins is None:
        nbins = int(max(h, w) / 2)
    nbins = max(1, int(nbins))

    y = np.arange(h) - cy
    x = np.arange(w) - cx
    X, Y = np.meshgrid(x, y)
    R = np.sqrt(X * X + Y * Y)

    Rmax = R.max()
    if Rmax <= 0:
        Rnorm = R
    else:
        Rnorm = R / (Rmax + 1e-12)
        Rnorm = np.minimum(Rnorm, 1.0 - 1e-12)

    bin_edges = np.linspace(0.0, 1.0, nbins + 1)
    bin_idx = np.digitize(Rnorm.ravel(), bin_edges) - 1
    bin_idx = np.clip(bin_idx, 0, nbins - 1)

    sums = np.bincount(bin_idx, weights=mag.ravel(), minlength=nbins)
    counts = np.bincount(bin_idx, minlength=nbins)

    radial_mean = np.zeros(nbins, dtype=np.float64)
    nonzero = counts > 0
    radial_mean[nonzero] = sums[nonzero] / counts[nonzero]

    bin_centers = 0.5 * (bin_edges[:-1] + bin_edges[1:])
    return bin_centers, radial_mean

def fourier_match_spectrum(img_arr: np.ndarray,
                           ref_img_arr: np.ndarray = None,
                           mode='auto',
                           alpha=1.0,
                           cutoff=0.25,
                           strength=0.9,
                           randomness=0.05,
                           phase_perturb=0.08,
                           radial_smooth=5,
                           seed=None):
    if seed is not None:
        rng = np.random.default_rng(seed)
    else:
        rng = np.random.default_rng()

    h, w = img_arr.shape[:2]
    cy, cx = h // 2, w // 2
    nbins = max(8, int(max(h, w) / 2))

    if mode == 'auto':
        mode = 'ref' if ref_img_arr is not None else 'model'

    bin_centers_src = np.linspace(0.0, 1.0, nbins)

    model_radial = None
    if mode == 'model':
        eps = 1e-8
        model_radial = (1.0 / (bin_centers_src + eps)) ** (alpha / 2.0)
        lf = max(1, nbins // 8)
        model_radial = model_radial / (np.median(model_radial[:lf]) + 1e-12)
        model_radial = gaussian_filter1d(model_radial, sigma=max(1, radial_smooth))

    ref_radial = None
    ref_bin_centers = None
    if mode == 'ref' and ref_img_arr is not None:
        if ref_img_arr.shape[0] != h or ref_img_arr.shape[1] != w:
            ref_img = Image.fromarray(ref_img_arr).resize((w, h), resample=Image.BICUBIC)
            ref_img_arr = np.array(ref_img)
        ref_gray = np.mean(ref_img_arr.astype(np.float32), axis=2) if ref_img_arr.ndim == 3 else ref_img_arr.astype(np.float32)
        Fref = np.fft.fftshift(np.fft.fft2(ref_gray))
        Mref = np.abs(Fref)
        ref_bin_centers, ref_radial = radial_profile(Mref, center=(h // 2, w // 2), nbins=nbins)
        ref_radial = gaussian_filter1d(ref_radial, sigma=max(1, radial_smooth))

    out = np.zeros_like(img_arr, dtype=np.float32)

    y = np.linspace(-1, 1, h, endpoint=False)[:, None]
    x = np.linspace(-1, 1, w, endpoint=False)[None, :]
    r = np.sqrt(x * x + y * y)
    r = np.clip(r, 0.0, 1.0 - 1e-6)

    for c in range(img_arr.shape[2]):
        channel = img_arr[:, :, c].astype(np.float32)
        F = np.fft.fft2(channel)
        Fshift = np.fft.fftshift(F)
        mag = np.abs(Fshift)
        phase = np.angle(Fshift)

        bin_centers_src_calc, src_radial = radial_profile(mag, center=(h // 2, w // 2), nbins=nbins)
        src_radial = gaussian_filter1d(src_radial, sigma=max(1, radial_smooth))
        bin_centers_src = bin_centers_src_calc

        if mode == 'ref' and ref_radial is not None:
            ref_interp = np.interp(bin_centers_src, ref_bin_centers, ref_radial)
            eps = 1e-8
            ratio = (ref_interp + eps) / (src_radial + eps)
            desired_radial = src_radial * ratio
        elif mode == 'model' and model_radial is not None:
            lf = max(1, nbins // 8)
            scale = (np.median(src_radial[:lf]) + 1e-12) / (np.median(model_radial[:lf]) + 1e-12)
            desired_radial = model_radial * scale
        else:
            desired_radial = src_radial.copy()

        eps = 1e-8
        multiplier_1d = (desired_radial + eps) / (src_radial + eps)
        multiplier_1d = np.clip(multiplier_1d, 0.2, 5.0)
        mult_2d = np.interp(r.ravel(), bin_centers_src, multiplier_1d).reshape(h, w)

        edge = 0.05 + 0.02 * (1.0 - cutoff) if 'cutoff' in globals() else 0.05
        edge = max(edge, 1e-6)
        weight = np.where(r <= 0.25, 1.0,
                         np.where(r <= 0.25 + edge,
                                  0.5 * (1 + np.cos(np.pi * (r - 0.25) / edge)),
                                  0.0))

        final_multiplier = 1.0 + (mult_2d - 1.0) * (weight * strength)

        if randomness and randomness > 0.0:
            noise = rng.normal(loc=1.0, scale=randomness, size=final_multiplier.shape)
            final_multiplier *= (1.0 + (noise - 1.0) * weight)

        mag2 = mag * final_multiplier

        if phase_perturb and phase_perturb > 0.0:
            phase_sigma = phase_perturb * np.clip((r - 0.25) / (1.0 - 0.25 + 1e-6), 0.0, 1.0)
            phase_noise = rng.standard_normal(size=phase_sigma.shape) * phase_sigma
            phase2 = phase + phase_noise
        else:
            phase2 = phase

        Fshift2 = mag2 * np.exp(1j * phase2)
        F_ishift = np.fft.ifftshift(Fshift2)
        img_back = np.fft.ifft2(F_ishift)
        img_back = np.real(img_back)

        blended = (1.0 - strength) * channel + strength * img_back
        out[:, :, c] = blended

    out = np.clip(out, 0, 255).astype(np.uint8)
    return out