File size: 11,877 Bytes
2366e36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Modified from https://github.com/clovaai/deep-text-recognition-benchmark
#
# Licensed under the Apache License, Version 2.0 (the "License");s
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://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.

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from mmocr.models.builder import PREPROCESSOR
from .base_preprocessor import BasePreprocessor


@PREPROCESSOR.register_module()
class TPSPreprocessor(BasePreprocessor):
    """Rectification Network of RARE, namely TPS based STN in
    https://arxiv.org/pdf/1603.03915.pdf.

    Args:
        num_fiducial (int): Number of fiducial points of TPS-STN.
        img_size (tuple(int, int)): Size :math:`(H, W)` of the input image.
        rectified_img_size (tuple(int, int)): Size :math:`(H_r, W_r)` of
            the rectified image.
        num_img_channel (int): Number of channels of the input image.
        init_cfg (dict or list[dict], optional): Initialization configs.
    """

    def __init__(self,
                 num_fiducial=20,
                 img_size=(32, 100),
                 rectified_img_size=(32, 100),
                 num_img_channel=1,
                 init_cfg=None):
        super().__init__(init_cfg=init_cfg)
        assert isinstance(num_fiducial, int)
        assert num_fiducial > 0
        assert isinstance(img_size, tuple)
        assert isinstance(rectified_img_size, tuple)
        assert isinstance(num_img_channel, int)

        self.num_fiducial = num_fiducial
        self.img_size = img_size
        self.rectified_img_size = rectified_img_size
        self.num_img_channel = num_img_channel
        self.LocalizationNetwork = LocalizationNetwork(self.num_fiducial,
                                                       self.num_img_channel)
        self.GridGenerator = GridGenerator(self.num_fiducial,
                                           self.rectified_img_size)

    def forward(self, batch_img):
        """
        Args:
            batch_img (Tensor): Images to be rectified with size
                :math:`(N, C, H, W)`.

        Returns:
            Tensor: Rectified image with size :math:`(N, C, H_r, W_r)`.
        """
        batch_C_prime = self.LocalizationNetwork(
            batch_img)  # batch_size x K x 2
        build_P_prime = self.GridGenerator.build_P_prime(
            batch_C_prime, batch_img.device
        )  # batch_size x n (= rectified_img_width x rectified_img_height) x 2
        build_P_prime_reshape = build_P_prime.reshape([
            build_P_prime.size(0), self.rectified_img_size[0],
            self.rectified_img_size[1], 2
        ])

        batch_rectified_img = F.grid_sample(
            batch_img,
            build_P_prime_reshape,
            padding_mode='border',
            align_corners=True)

        return batch_rectified_img


class LocalizationNetwork(nn.Module):
    """Localization Network of RARE, which predicts C' (K x 2) from input
    (img_width x img_height)

    Args:
        num_fiducial (int): Number of fiducial points of TPS-STN.
        num_img_channel (int): Number of channels of the input image.
    """

    def __init__(self, num_fiducial, num_img_channel):
        super().__init__()
        self.num_fiducial = num_fiducial
        self.num_img_channel = num_img_channel
        self.conv = nn.Sequential(
            nn.Conv2d(
                in_channels=self.num_img_channel,
                out_channels=64,
                kernel_size=3,
                stride=1,
                padding=1,
                bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(True),
            nn.MaxPool2d(2, 2),  # batch_size x 64 x img_height/2 x img_width/2
            nn.Conv2d(64, 128, 3, 1, 1, bias=False),
            nn.BatchNorm2d(128),
            nn.ReLU(True),
            nn.MaxPool2d(2, 2),  # batch_size x 128 x img_h/4 x img_w/4
            nn.Conv2d(128, 256, 3, 1, 1, bias=False),
            nn.BatchNorm2d(256),
            nn.ReLU(True),
            nn.MaxPool2d(2, 2),  # batch_size x 256 x img_h/8 x img_w/8
            nn.Conv2d(256, 512, 3, 1, 1, bias=False),
            nn.BatchNorm2d(512),
            nn.ReLU(True),
            nn.AdaptiveAvgPool2d(1)  # batch_size x 512
        )

        self.localization_fc1 = nn.Sequential(
            nn.Linear(512, 256), nn.ReLU(True))
        self.localization_fc2 = nn.Linear(256, self.num_fiducial * 2)

        # Init fc2 in LocalizationNetwork
        self.localization_fc2.weight.data.fill_(0)
        ctrl_pts_x = np.linspace(-1.0, 1.0, int(num_fiducial / 2))
        ctrl_pts_y_top = np.linspace(0.0, -1.0, num=int(num_fiducial / 2))
        ctrl_pts_y_bottom = np.linspace(1.0, 0.0, num=int(num_fiducial / 2))
        ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
        ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
        initial_bias = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0)
        self.localization_fc2.bias.data = torch.from_numpy(
            initial_bias).float().view(-1)

    def forward(self, batch_img):
        """
        Args:
            batch_img (Tensor): Batch input image of shape
                :math:`(N, C, H, W)`.

        Returns:
            Tensor: Predicted coordinates of fiducial points for input batch.
            The shape is :math:`(N, F, 2)` where :math:`F` is ``num_fiducial``.
        """
        batch_size = batch_img.size(0)
        features = self.conv(batch_img).view(batch_size, -1)
        batch_C_prime = self.localization_fc2(
            self.localization_fc1(features)).view(batch_size,
                                                  self.num_fiducial, 2)
        return batch_C_prime


class GridGenerator(nn.Module):
    """Grid Generator of RARE, which produces P_prime by multiplying T with P.

    Args:
        num_fiducial (int): Number of fiducial points of TPS-STN.
        rectified_img_size (tuple(int, int)):
            Size :math:`(H_r, W_r)` of the rectified image.
    """

    def __init__(self, num_fiducial, rectified_img_size):
        """Generate P_hat and inv_delta_C for later."""
        super().__init__()
        self.eps = 1e-6
        self.rectified_img_height = rectified_img_size[0]
        self.rectified_img_width = rectified_img_size[1]
        self.num_fiducial = num_fiducial
        self.C = self._build_C(self.num_fiducial)  # num_fiducial x 2
        self.P = self._build_P(self.rectified_img_width,
                               self.rectified_img_height)
        # for multi-gpu, you need register buffer
        self.register_buffer(
            'inv_delta_C',
            torch.tensor(self._build_inv_delta_C(
                self.num_fiducial,
                self.C)).float())  # num_fiducial+3 x num_fiducial+3
        self.register_buffer('P_hat',
                             torch.tensor(
                                 self._build_P_hat(
                                     self.num_fiducial, self.C,
                                     self.P)).float())  # n x num_fiducial+3
        # for fine-tuning with different image width,
        # you may use below instead of self.register_buffer
        # self.inv_delta_C = torch.tensor(
        #     self._build_inv_delta_C(
        #         self.num_fiducial,
        #         self.C)).float().cuda()  # num_fiducial+3 x num_fiducial+3
        # self.P_hat = torch.tensor(
        #     self._build_P_hat(self.num_fiducial, self.C,
        #                       self.P)).float().cuda()  # n x num_fiducial+3

    def _build_C(self, num_fiducial):
        """Return coordinates of fiducial points in rectified_img; C."""
        ctrl_pts_x = np.linspace(-1.0, 1.0, int(num_fiducial / 2))
        ctrl_pts_y_top = -1 * np.ones(int(num_fiducial / 2))
        ctrl_pts_y_bottom = np.ones(int(num_fiducial / 2))
        ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
        ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
        C = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0)
        return C  # num_fiducial x 2

    def _build_inv_delta_C(self, num_fiducial, C):
        """Return inv_delta_C which is needed to calculate T."""
        hat_C = np.zeros((num_fiducial, num_fiducial), dtype=float)
        for i in range(0, num_fiducial):
            for j in range(i, num_fiducial):
                r = np.linalg.norm(C[i] - C[j])
                hat_C[i, j] = r
                hat_C[j, i] = r
        np.fill_diagonal(hat_C, 1)
        hat_C = (hat_C**2) * np.log(hat_C)
        # print(C.shape, hat_C.shape)
        delta_C = np.concatenate(  # num_fiducial+3 x num_fiducial+3
            [
                np.concatenate([np.ones((num_fiducial, 1)), C, hat_C],
                               axis=1),  # num_fiducial x num_fiducial+3
                np.concatenate([np.zeros(
                    (2, 3)), np.transpose(C)], axis=1),  # 2 x num_fiducial+3
                np.concatenate([np.zeros(
                    (1, 3)), np.ones((1, num_fiducial))],
                               axis=1)  # 1 x num_fiducial+3
            ],
            axis=0)
        inv_delta_C = np.linalg.inv(delta_C)
        return inv_delta_C  # num_fiducial+3 x num_fiducial+3

    def _build_P(self, rectified_img_width, rectified_img_height):
        rectified_img_grid_x = (
            np.arange(-rectified_img_width, rectified_img_width, 2) +
            1.0) / rectified_img_width  # self.rectified_img_width
        rectified_img_grid_y = (
            np.arange(-rectified_img_height, rectified_img_height, 2) +
            1.0) / rectified_img_height  # self.rectified_img_height
        P = np.stack(  # self.rectified_img_w x self.rectified_img_h x 2
            np.meshgrid(rectified_img_grid_x, rectified_img_grid_y),
            axis=2)
        return P.reshape([
            -1, 2
        ])  # n (= self.rectified_img_width x self.rectified_img_height) x 2

    def _build_P_hat(self, num_fiducial, C, P):
        n = P.shape[
            0]  # n (= self.rectified_img_width x self.rectified_img_height)
        P_tile = np.tile(np.expand_dims(P, axis=1),
                         (1, num_fiducial,
                          1))  # n x 2 -> n x 1 x 2 -> n x num_fiducial x 2
        C_tile = np.expand_dims(C, axis=0)  # 1 x num_fiducial x 2
        P_diff = P_tile - C_tile  # n x num_fiducial x 2
        rbf_norm = np.linalg.norm(
            P_diff, ord=2, axis=2, keepdims=False)  # n x num_fiducial
        rbf = np.multiply(np.square(rbf_norm),
                          np.log(rbf_norm + self.eps))  # n x num_fiducial
        P_hat = np.concatenate([np.ones((n, 1)), P, rbf], axis=1)
        return P_hat  # n x num_fiducial+3

    def build_P_prime(self, batch_C_prime, device='cuda'):
        """Generate Grid from batch_C_prime [batch_size x num_fiducial x 2]"""
        batch_size = batch_C_prime.size(0)
        batch_inv_delta_C = self.inv_delta_C.repeat(batch_size, 1, 1)
        batch_P_hat = self.P_hat.repeat(batch_size, 1, 1)
        batch_C_prime_with_zeros = torch.cat(
            (batch_C_prime, torch.zeros(batch_size, 3, 2).float().to(device)),
            dim=1)  # batch_size x num_fiducial+3 x 2
        batch_T = torch.bmm(
            batch_inv_delta_C,
            batch_C_prime_with_zeros)  # batch_size x num_fiducial+3 x 2
        batch_P_prime = torch.bmm(batch_P_hat, batch_T)  # batch_size x n x 2
        return batch_P_prime  # batch_size x n x 2