File size: 4,889 Bytes
dadf189
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Joint augmentation for fingerprint images and minutiae.

Applies the same geometric transform to both the image and minutiae
coordinates simultaneously, ensuring spatial consistency.

Augmentations:
    - Random rotation: rotates image + rotation matrix on (x,y) + adds angle to θ
    - Random translation: shifts image + adds offset to (x,y)
    - Minutia dropout: randomly drops points (keeps ≥ min_keep)
    - Coordinate jitter: Gaussian noise on (x,y) — minutiae only
"""

import math
import random

import torch
import torch.nn.functional as F

from ..configs.default import AugmentConfig


class JointAugmentor:
    """Applies the same geometric transform to image and minutiae simultaneously.

    Parameters
    ----------
    cfg : AugmentConfig
        Augmentation configuration.
    """

    def __init__(self, cfg: AugmentConfig):
        self.cfg = cfg

    def __call__(
        self,
        image: torch.Tensor,
        minutiae: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            image:    ``(1, H, W)`` normalized [0, 1] fingerprint image.
            minutiae: ``(N, 3)`` raw ``[x, y, θ]`` in pixel coordinates.

        Returns:
            image:    ``(1, H, W)`` augmented image.
            minutiae: ``(N', 3)`` augmented minutiae (N' ≤ N after dropout).
        """
        _, H, W = image.shape
        m = minutiae.clone()

        # Center of image for rotation
        cx, cy = (W - 1) / 2.0, (H - 1) / 2.0

        # 1. Random rotation
        if self.cfg.rotate:
            angle_deg = random.uniform(-self.cfg.rotate_range, self.cfg.rotate_range)
            angle_rad = math.radians(angle_deg)
            cos_a, sin_a = math.cos(angle_rad), math.sin(angle_rad)

            # Rotate image using affine grid
            # Rotation matrix (clockwise in pixel coords)
            theta = torch.tensor(
                [
                    [cos_a, -sin_a, 0.0],
                    [sin_a, cos_a, 0.0],
                ],
                dtype=image.dtype,
            ).unsqueeze(0)  # (1, 2, 3)

            grid = F.affine_grid(theta, [1, 1, H, W], align_corners=True)
            image = F.grid_sample(
                image.unsqueeze(0),
                grid,
                mode="bilinear",
                padding_mode="zeros",
                align_corners=True,
            ).squeeze(0)

            # Rotate minutiae coordinates around image center
            x_centered = m[:, 0] - cx
            y_centered = m[:, 1] - cy
            m[:, 0] = cos_a * x_centered - sin_a * y_centered + cx
            m[:, 1] = sin_a * x_centered + cos_a * y_centered + cy

            # Rotate orientation
            m[:, 2] = m[:, 2] + angle_rad
            m[:, 2] = torch.atan2(torch.sin(m[:, 2]), torch.cos(m[:, 2]))

        # 2. Random translation
        if self.cfg.translate and self.cfg.translate_range > 0:
            tx = random.uniform(-self.cfg.translate_range, self.cfg.translate_range)
            ty = random.uniform(-self.cfg.translate_range, self.cfg.translate_range)

            # Translate image using affine grid
            theta = torch.tensor(
                [
                    [1.0, 0.0, -2.0 * tx / (W - 1)],
                    [0.0, 1.0, -2.0 * ty / (H - 1)],
                ],
                dtype=image.dtype,
            ).unsqueeze(0)

            grid = F.affine_grid(theta, [1, 1, H, W], align_corners=True)
            image = F.grid_sample(
                image.unsqueeze(0),
                grid,
                mode="bilinear",
                padding_mode="zeros",
                align_corners=True,
            ).squeeze(0)

            # Translate minutiae
            m[:, 0] += tx
            m[:, 1] += ty

        # 3. Minutia dropout
        if self.cfg.minutia_dropout > 0 and m.shape[0] > self.cfg.min_keep:
            keep = torch.rand(m.shape[0]) > self.cfg.minutia_dropout
            if keep.sum() < self.cfg.min_keep:
                keep[: self.cfg.min_keep] = True
            m = m[keep]

        # 4. Coordinate jitter (minutiae only, image unchanged)
        if self.cfg.jitter_std > 0:
            noise = torch.randn(m.shape[0], 2) * self.cfg.jitter_std
            m[:, :2] += noise

        # 5. Spurious minutiae insertion
        spurious_rate = getattr(self.cfg, "spurious_rate", 0.0)
        if spurious_rate > 0 and m.shape[0] > 0:
            n_spurious = max(1, int(m.shape[0] * spurious_rate))
            xy_min = m[:, :2].min(dim=0).values
            xy_max = m[:, :2].max(dim=0).values
            xy_range = (xy_max - xy_min).clamp(min=1.0)
            fake_xy = xy_min + torch.rand(n_spurious, 2) * xy_range
            fake_theta = torch.rand(n_spurious, 1) * 2 * math.pi - math.pi
            fake = torch.cat([fake_xy, fake_theta], dim=-1)
            m = torch.cat([m, fake], dim=0)

        return image, m