File size: 4,712 Bytes
29f689c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Scene Text Recognition Model Hub
# Copyright 2022 Darwin Bautista
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from functools import partial

import imgaug.augmenters as iaa
import numpy as np
from PIL import Image, ImageFilter

from openrec.preprocess import auto_augment
from openrec.preprocess.auto_augment import _LEVEL_DENOM, LEVEL_TO_ARG, NAME_TO_OP, _randomly_negate, rotate


def rotate_expand(img, degrees, **kwargs):
    """Rotate operation with expand=True to avoid cutting off the
    characters."""
    kwargs['expand'] = True
    return rotate(img, degrees, **kwargs)


def _level_to_arg(level, hparams, key, default):
    magnitude = hparams.get(key, default)
    level = (level / _LEVEL_DENOM) * magnitude
    level = _randomly_negate(level)
    return level,


def apply():
    # Overrides
    NAME_TO_OP.update({'Rotate': rotate_expand})
    LEVEL_TO_ARG.update({
        'Rotate':
        partial(_level_to_arg, key='rotate_deg', default=30.),
        'ShearX':
        partial(_level_to_arg, key='shear_x_pct', default=0.3),
        'ShearY':
        partial(_level_to_arg, key='shear_y_pct', default=0.3),
        'TranslateXRel':
        partial(_level_to_arg, key='translate_x_pct', default=0.45),
        'TranslateYRel':
        partial(_level_to_arg, key='translate_y_pct', default=0.45),
    })


apply()

_OP_CACHE = {}


def _get_op(key, factory):
    try:
        op = _OP_CACHE[key]
    except KeyError:
        op = factory()
        _OP_CACHE[key] = op
    return op


def _get_param(level, img, max_dim_factor, min_level=1):
    max_level = max(min_level, max_dim_factor * max(img.size))
    return round(min(level, max_level))


def gaussian_blur(img, radius, **__):
    radius = _get_param(radius, img, 0.02)
    key = 'gaussian_blur_' + str(radius)
    op = _get_op(key, lambda: ImageFilter.GaussianBlur(radius))
    return img.filter(op)


def motion_blur(img, k, **__):
    k = _get_param(k, img, 0.08, 3) | 1  # bin to odd values
    key = 'motion_blur_' + str(k)
    op = _get_op(key, lambda: iaa.MotionBlur(k))
    return Image.fromarray(op(image=np.asarray(img)))


def gaussian_noise(img, scale, **_):
    scale = _get_param(scale, img, 0.25) | 1  # bin to odd values
    key = 'gaussian_noise_' + str(scale)
    op = _get_op(key, lambda: iaa.AdditiveGaussianNoise(scale=scale))
    return Image.fromarray(op(image=np.asarray(img)))


def poisson_noise(img, lam, **_):
    lam = _get_param(lam, img, 0.2) | 1  # bin to odd values
    key = 'poisson_noise_' + str(lam)
    op = _get_op(key, lambda: iaa.AdditivePoissonNoise(lam))
    return Image.fromarray(op(image=np.asarray(img)))


def _level_to_arg(level, _hparams, max):
    level = max * level / auto_augment._LEVEL_DENOM
    return level,


_RAND_TRANSFORMS = auto_augment._RAND_INCREASING_TRANSFORMS.copy()
_RAND_TRANSFORMS.remove(
    'SharpnessIncreasing')  # remove, interferes with *blur ops
_RAND_TRANSFORMS.extend([
    'GaussianBlur',
    # 'MotionBlur',
    # 'GaussianNoise',
    'PoissonNoise'
])
auto_augment.LEVEL_TO_ARG.update({
    'GaussianBlur':
    partial(_level_to_arg, max=4),
    'MotionBlur':
    partial(_level_to_arg, max=20),
    'GaussianNoise':
    partial(_level_to_arg, max=0.1 * 255),
    'PoissonNoise':
    partial(_level_to_arg, max=40)
})
auto_augment.NAME_TO_OP.update({
    'GaussianBlur': gaussian_blur,
    'MotionBlur': motion_blur,
    'GaussianNoise': gaussian_noise,
    'PoissonNoise': poisson_noise
})


def rand_augment_transform(magnitude=5, num_layers=3):
    # These are tuned for magnitude=5, which means that effective magnitudes are half of these values.
    hparams = {
        'rotate_deg': 30,
        'shear_x_pct': 0.9,
        'shear_y_pct': 0.2,
        'translate_x_pct': 0.10,
        'translate_y_pct': 0.30
    }
    ra_ops = auto_augment.rand_augment_ops(magnitude,
                                           hparams=hparams,
                                           transforms=_RAND_TRANSFORMS)
    # Supply weights to disable replacement in random selection (i.e. avoid applying the same op twice)
    choice_weights = [1. / len(ra_ops) for _ in range(len(ra_ops))]
    return auto_augment.RandAugment(ra_ops, num_layers, choice_weights)