Merge branch 'dygraph' into pgnet-readme

ppocr/modeling/heads/__init__.py

@@ -20,6 +20,7 @@ def build_head(config):
     from .det_db_head import DBHead
     from .det_east_head import EASTHead
     from .det_sast_head import SASTHead
+    from .e2e_pg_head import PGHead
 
     # rec head
     from .rec_ctc_head import CTCHead
@@ -30,8 +31,8 @@ def build_head(config):
     from .cls_head import ClsHead
     support_dict = [
         'DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead',
-        'SRNHead'
-    ]
+        'SRNHead', 'PGHead']
+
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('head only support {}'.format(

ppocr/modeling/heads/e2e_pg_head.py

+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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
+#
+#    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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 padding,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            param_attr=ParamAttr(name="bn_" + name + "_scale"),
+            bias_attr=ParamAttr(name="bn_" + name + "_offset"),
+            moving_mean_name="bn_" + name + "_mean",
+            moving_variance_name="bn_" + name + "_variance",
+            use_global_stats=False)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class PGHead(nn.Layer):
+    """
+    """
+
+    def __init__(self, in_channels, **kwargs):
+        super(PGHead, self).__init__()
+        self.conv_f_score1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=64,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_score{}".format(1))
+        self.conv_f_score2 = ConvBNLayer(
+            in_channels=64,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            act='relu',
+            name="conv_f_score{}".format(2))
+        self.conv_f_score3 = ConvBNLayer(
+            in_channels=64,
+            out_channels=128,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_score{}".format(3))
+
+        self.conv1 = nn.Conv2D(
+            in_channels=128,
+            out_channels=1,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            groups=1,
+            weight_attr=ParamAttr(name="conv_f_score{}".format(4)),
+            bias_attr=False)
+
+        self.conv_f_boder1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=64,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_boder{}".format(1))
+        self.conv_f_boder2 = ConvBNLayer(
+            in_channels=64,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            act='relu',
+            name="conv_f_boder{}".format(2))
+        self.conv_f_boder3 = ConvBNLayer(
+            in_channels=64,
+            out_channels=128,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_boder{}".format(3))
+        self.conv2 = nn.Conv2D(
+            in_channels=128,
+            out_channels=4,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            groups=1,
+            weight_attr=ParamAttr(name="conv_f_boder{}".format(4)),
+            bias_attr=False)
+        self.conv_f_char1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=128,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_char{}".format(1))
+        self.conv_f_char2 = ConvBNLayer(
+            in_channels=128,
+            out_channels=128,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            act='relu',
+            name="conv_f_char{}".format(2))
+        self.conv_f_char3 = ConvBNLayer(
+            in_channels=128,
+            out_channels=256,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_char{}".format(3))
+        self.conv_f_char4 = ConvBNLayer(
+            in_channels=256,
+            out_channels=256,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            act='relu',
+            name="conv_f_char{}".format(4))
+        self.conv_f_char5 = ConvBNLayer(
+            in_channels=256,
+            out_channels=256,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_char{}".format(5))
+        self.conv3 = nn.Conv2D(
+            in_channels=256,
+            out_channels=37,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            groups=1,
+            weight_attr=ParamAttr(name="conv_f_char{}".format(6)),
+            bias_attr=False)
+
+        self.conv_f_direc1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=64,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_direc{}".format(1))
+        self.conv_f_direc2 = ConvBNLayer(
+            in_channels=64,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            act='relu',
+            name="conv_f_direc{}".format(2))
+        self.conv_f_direc3 = ConvBNLayer(
+            in_channels=64,
+            out_channels=128,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            act='relu',
+            name="conv_f_direc{}".format(3))
+        self.conv4 = nn.Conv2D(
+            in_channels=128,
+            out_channels=2,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            groups=1,
+            weight_attr=ParamAttr(name="conv_f_direc{}".format(4)),
+            bias_attr=False)
+
+    def forward(self, x):
+        f_score = self.conv_f_score1(x)
+        f_score = self.conv_f_score2(f_score)
+        f_score = self.conv_f_score3(f_score)
+        f_score = self.conv1(f_score)
+        f_score = F.sigmoid(f_score)
+
+        # f_border
+        f_border = self.conv_f_boder1(x)
+        f_border = self.conv_f_boder2(f_border)
+        f_border = self.conv_f_boder3(f_border)
+        f_border = self.conv2(f_border)
+
+        f_char = self.conv_f_char1(x)
+        f_char = self.conv_f_char2(f_char)
+        f_char = self.conv_f_char3(f_char)
+        f_char = self.conv_f_char4(f_char)
+        f_char = self.conv_f_char5(f_char)
+        f_char = self.conv3(f_char)
+
+        f_direction = self.conv_f_direc1(x)
+        f_direction = self.conv_f_direc2(f_direction)
+        f_direction = self.conv_f_direc3(f_direction)
+        f_direction = self.conv4(f_direction)
+
+        predicts = {}
+        predicts['f_score'] = f_score
+        predicts['f_border'] = f_border
+        predicts['f_char'] = f_char
+        predicts['f_direction'] = f_direction
+        return predicts

ppocr/modeling/necks/__init__.py

@@ -14,12 +14,14 @@
 
 __all__ = ['build_neck']
 
+
 def build_neck(config):
     from .db_fpn import DBFPN
     from .east_fpn import EASTFPN
     from .sast_fpn import SASTFPN
     from .rnn import SequenceEncoder
-    support_dict = ['DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder']
+    from .pg_fpn import PGFPN
+    support_dict = ['DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder', 'PGFPN']
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('neck only support {}'.format(

ppocr/modeling/necks/pg_fpn.py

+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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
+#
+#    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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 groups=1,
+                 is_vd_mode=False,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+
+        self.is_vd_mode = is_vd_mode
+        self._pool2d_avg = nn.AvgPool2D(
+            kernel_size=2, stride=2, padding=0, ceil_mode=True)
+        self._conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + "_weights"),
+            bias_attr=False)
+        if name == "conv1":
+            bn_name = "bn_" + name
+        else:
+            bn_name = "bn" + name[3:]
+        self._batch_norm = nn.BatchNorm(
+            out_channels,
+            act=act,
+            param_attr=ParamAttr(name=bn_name + '_scale'),
+            bias_attr=ParamAttr(bn_name + '_offset'),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance',
+            use_global_stats=False)
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        return y
+
+
+class DeConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=4,
+                 stride=2,
+                 padding=1,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(DeConvBNLayer, self).__init__()
+
+        self.if_act = if_act
+        self.act = act
+        self.deconv = nn.Conv2DTranspose(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            param_attr=ParamAttr(name="bn_" + name + "_scale"),
+            bias_attr=ParamAttr(name="bn_" + name + "_offset"),
+            moving_mean_name="bn_" + name + "_mean",
+            moving_variance_name="bn_" + name + "_variance",
+            use_global_stats=False)
+
+    def forward(self, x):
+        x = self.deconv(x)
+        x = self.bn(x)
+        return x
+
+
+class PGFPN(nn.Layer):
+    def __init__(self, in_channels, **kwargs):
+        super(PGFPN, self).__init__()
+        num_inputs = [2048, 2048, 1024, 512, 256]
+        num_outputs = [256, 256, 192, 192, 128]
+        self.out_channels = 128
+        self.conv_bn_layer_1 = ConvBNLayer(
+            in_channels=3,
+            out_channels=32,
+            kernel_size=3,
+            stride=1,
+            act=None,
+            name='FPN_d1')
+        self.conv_bn_layer_2 = ConvBNLayer(
+            in_channels=64,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            act=None,
+            name='FPN_d2')
+        self.conv_bn_layer_3 = ConvBNLayer(
+            in_channels=256,
+            out_channels=128,
+            kernel_size=3,
+            stride=1,
+            act=None,
+            name='FPN_d3')
+        self.conv_bn_layer_4 = ConvBNLayer(
+            in_channels=32,
+            out_channels=64,
+            kernel_size=3,
+            stride=2,
+            act=None,
+            name='FPN_d4')
+        self.conv_bn_layer_5 = ConvBNLayer(
+            in_channels=64,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name='FPN_d5')
+        self.conv_bn_layer_6 = ConvBNLayer(
+            in_channels=64,
+            out_channels=128,
+            kernel_size=3,
+            stride=2,
+            act=None,
+            name='FPN_d6')
+        self.conv_bn_layer_7 = ConvBNLayer(
+            in_channels=128,
+            out_channels=128,
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name='FPN_d7')
+        self.conv_bn_layer_8 = ConvBNLayer(
+            in_channels=128,
+            out_channels=128,
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name='FPN_d8')
+
+        self.conv_h0 = ConvBNLayer(
+            in_channels=num_inputs[0],
+            out_channels=num_outputs[0],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_h{}".format(0))
+        self.conv_h1 = ConvBNLayer(
+            in_channels=num_inputs[1],
+            out_channels=num_outputs[1],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_h{}".format(1))
+        self.conv_h2 = ConvBNLayer(
+            in_channels=num_inputs[2],
+            out_channels=num_outputs[2],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_h{}".format(2))
+        self.conv_h3 = ConvBNLayer(
+            in_channels=num_inputs[3],
+            out_channels=num_outputs[3],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_h{}".format(3))
+        self.conv_h4 = ConvBNLayer(
+            in_channels=num_inputs[4],
+            out_channels=num_outputs[4],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_h{}".format(4))
+
+        self.dconv0 = DeConvBNLayer(
+            in_channels=num_outputs[0],
+            out_channels=num_outputs[0 + 1],
+            name="dconv_{}".format(0))
+        self.dconv1 = DeConvBNLayer(
+            in_channels=num_outputs[1],
+            out_channels=num_outputs[1 + 1],
+            act=None,
+            name="dconv_{}".format(1))
+        self.dconv2 = DeConvBNLayer(
+            in_channels=num_outputs[2],
+            out_channels=num_outputs[2 + 1],
+            act=None,
+            name="dconv_{}".format(2))
+        self.dconv3 = DeConvBNLayer(
+            in_channels=num_outputs[3],
+            out_channels=num_outputs[3 + 1],
+            act=None,
+            name="dconv_{}".format(3))
+        self.conv_g1 = ConvBNLayer(
+            in_channels=num_outputs[1],
+            out_channels=num_outputs[1],
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv_g{}".format(1))
+        self.conv_g2 = ConvBNLayer(
+            in_channels=num_outputs[2],
+            out_channels=num_outputs[2],
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv_g{}".format(2))
+        self.conv_g3 = ConvBNLayer(
+            in_channels=num_outputs[3],
+            out_channels=num_outputs[3],
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv_g{}".format(3))
+        self.conv_g4 = ConvBNLayer(
+            in_channels=num_outputs[4],
+            out_channels=num_outputs[4],
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv_g{}".format(4))
+        self.convf = ConvBNLayer(
+            in_channels=num_outputs[4],
+            out_channels=num_outputs[4],
+            kernel_size=1,
+            stride=1,
+            act=None,
+            name="conv_f{}".format(4))
+
+    def forward(self, x):
+        c0, c1, c2, c3, c4, c5, c6 = x
+        # FPN_Down_Fusion
+        f = [c0, c1, c2]
+        g = [None, None, None]
+        h = [None, None, None]
+        h[0] = self.conv_bn_layer_1(f[0])
+        h[1] = self.conv_bn_layer_2(f[1])
+        h[2] = self.conv_bn_layer_3(f[2])
+
+        g[0] = self.conv_bn_layer_4(h[0])
+        g[1] = paddle.add(g[0], h[1])
+        g[1] = F.relu(g[1])
+        g[1] = self.conv_bn_layer_5(g[1])
+        g[1] = self.conv_bn_layer_6(g[1])
+
+        g[2] = paddle.add(g[1], h[2])
+        g[2] = F.relu(g[2])
+        g[2] = self.conv_bn_layer_7(g[2])
+        f_down = self.conv_bn_layer_8(g[2])
+
+        # FPN UP Fusion
+        f1 = [c6, c5, c4, c3, c2]
+        g = [None, None, None, None, None]
+        h = [None, None, None, None, None]
+        h[0] = self.conv_h0(f1[0])
+        h[1] = self.conv_h1(f1[1])
+        h[2] = self.conv_h2(f1[2])
+        h[3] = self.conv_h3(f1[3])
+        h[4] = self.conv_h4(f1[4])
+
+        g[0] = self.dconv0(h[0])
+        g[1] = paddle.add(g[0], h[1])
+        g[1] = F.relu(g[1])
+        g[1] = self.conv_g1(g[1])
+        g[1] = self.dconv1(g[1])
+
+        g[2] = paddle.add(g[1], h[2])
+        g[2] = F.relu(g[2])
+        g[2] = self.conv_g2(g[2])
+        g[2] = self.dconv2(g[2])
+
+        g[3] = paddle.add(g[2], h[3])
+        g[3] = F.relu(g[3])
+        g[3] = self.conv_g3(g[3])
+        g[3] = self.dconv3(g[3])
+
+        g[4] = paddle.add(x=g[3], y=h[4])
+        g[4] = F.relu(g[4])
+        g[4] = self.conv_g4(g[4])
+        f_up = self.convf(g[4])
+        f_common = paddle.add(f_down, f_up)
+        f_common = F.relu(f_common)
+        return f_common

ppocr/postprocess/__init__.py

@@ -28,10 +28,11 @@ def build_post_process(config, global_config=None):
     from .sast_postprocess import SASTPostProcess
     from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode
     from .cls_postprocess import ClsPostProcess
+    from .pg_postprocess import PGPostProcess
 
     support_dict = [
         'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode',
-        'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode'
+        'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode', 'PGPostProcess'
     ]
 
     config = copy.deepcopy(config)

ppocr/postprocess/pg_postprocess.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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
+#
+#     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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import sys
+
+__dir__ = os.path.dirname(__file__)
+sys.path.append(__dir__)
+sys.path.append(os.path.join(__dir__, '..'))
+
+from ppocr.utils.e2e_utils.extract_textpoint import *
+from ppocr.utils.e2e_utils.visual import *
+import paddle
+
+
+class PGPostProcess(object):
+    """
+    The post process for PGNet.
+    """
+
+    def __init__(self, character_dict_path, valid_set, score_thresh, **kwargs):
+        self.Lexicon_Table = get_dict(character_dict_path)
+        self.valid_set = valid_set
+        self.score_thresh = score_thresh
+
+        # c++ la-nms is faster, but only support python 3.5
+        self.is_python35 = False
+        if sys.version_info.major == 3 and sys.version_info.minor == 5:
+            self.is_python35 = True
+
+    def __call__(self, outs_dict, shape_list):
+        p_score = outs_dict['f_score']
+        p_border = outs_dict['f_border']
+        p_char = outs_dict['f_char']
+        p_direction = outs_dict['f_direction']
+        if isinstance(p_score, paddle.Tensor):
+            p_score = p_score[0].numpy()
+            p_border = p_border[0].numpy()
+            p_direction = p_direction[0].numpy()
+            p_char = p_char[0].numpy()
+        else:
+            p_score = p_score[0]
+            p_border = p_border[0]
+            p_direction = p_direction[0]
+            p_char = p_char[0]
+        src_h, src_w, ratio_h, ratio_w = shape_list[0]
+        is_curved = self.valid_set == "totaltext"
+        instance_yxs_list = generate_pivot_list(
+            p_score,
+            p_char,
+            p_direction,
+            score_thresh=self.score_thresh,
+            is_backbone=True,
+            is_curved=is_curved)
+        p_char = paddle.to_tensor(np.expand_dims(p_char, axis=0))
+        char_seq_idx_set = []
+        for i in range(len(instance_yxs_list)):
+            gather_info_lod = paddle.to_tensor(instance_yxs_list[i])
+            f_char_map = paddle.transpose(p_char, [0, 2, 3, 1])
+            feature_seq = paddle.gather_nd(f_char_map, gather_info_lod)
+            feature_seq = np.expand_dims(feature_seq.numpy(), axis=0)
+            feature_len = [len(feature_seq[0])]
+            featyre_seq = paddle.to_tensor(feature_seq)
+            feature_len = np.array([feature_len]).astype(np.int64)
+            length = paddle.to_tensor(feature_len)
+            seq_pred = paddle.fluid.layers.ctc_greedy_decoder(
+                input=featyre_seq, blank=36, input_length=length)
+            seq_pred_str = seq_pred[0].numpy().tolist()[0]
+            seq_len = seq_pred[1].numpy()[0][0]
+            temp_t = []
+            for c in seq_pred_str[:seq_len]:
+                temp_t.append(c)
+            char_seq_idx_set.append(temp_t)
+        seq_strs = []
+        for char_idx_set in char_seq_idx_set:
+            pr_str = ''.join([self.Lexicon_Table[pos] for pos in char_idx_set])
+            seq_strs.append(pr_str)
+        poly_list = []
+        keep_str_list = []
+        all_point_list = []
+        all_point_pair_list = []
+        for yx_center_line, keep_str in zip(instance_yxs_list, seq_strs):
+            if len(yx_center_line) == 1:
+                yx_center_line.append(yx_center_line[-1])
+
+            offset_expand = 1.0
+            if self.valid_set == 'totaltext':
+                offset_expand = 1.2
+
+            point_pair_list = []
+            for batch_id, y, x in yx_center_line:
+                offset = p_border[:, y, x].reshape(2, 2)
+                if offset_expand != 1.0:
+                    offset_length = np.linalg.norm(
+                        offset, axis=1, keepdims=True)
+                    expand_length = np.clip(
+                        offset_length * (offset_expand - 1),
+                        a_min=0.5,
+                        a_max=3.0)
+                    offset_detal = offset / offset_length * expand_length
+                    offset = offset + offset_detal
+                ori_yx = np.array([y, x], dtype=np.float32)
+                point_pair = (ori_yx + offset)[:, ::-1] * 4.0 / np.array(
+                    [ratio_w, ratio_h]).reshape(-1, 2)
+                point_pair_list.append(point_pair)
+
+                all_point_list.append([
+                    int(round(x * 4.0 / ratio_w)),
+                    int(round(y * 4.0 / ratio_h))
+                ])
+                all_point_pair_list.append(point_pair.round().astype(np.int32)
+                                           .tolist())
+
+            detected_poly, pair_length_info = point_pair2poly(point_pair_list)
+            detected_poly = expand_poly_along_width(
+                detected_poly, shrink_ratio_of_width=0.2)
+            detected_poly[:, 0] = np.clip(
+                detected_poly[:, 0], a_min=0, a_max=src_w)
+            detected_poly[:, 1] = np.clip(
+                detected_poly[:, 1], a_min=0, a_max=src_h)
+
+            if len(keep_str) < 2:
+                continue
+
+            keep_str_list.append(keep_str)
+            if self.valid_set == 'partvgg':
+                middle_point = len(detected_poly) // 2
+                detected_poly = detected_poly[
+                    [0, middle_point - 1, middle_point, -1], :]
+                poly_list.append(detected_poly)
+            elif self.valid_set == 'totaltext':
+                poly_list.append(detected_poly)
+            else:
+                print('--> Not supported format.')
+                exit(-1)
+        data = {
+            'points': poly_list,
+            'strs': keep_str_list,
+        }
+        return data

ppocr/postprocess/sast_postprocess.py

@@ -18,6 +18,7 @@ from __future__ import print_function
 
 import os
 import sys
+
 __dir__ = os.path.dirname(__file__)
 sys.path.append(__dir__)
 sys.path.append(os.path.join(__dir__, '..'))
@@ -67,11 +68,15 @@ class SASTPostProcess(object):
             point_list[point_num - 1 - idx] = point_pair[1]
         return np.array(point_list).reshape(-1, 2)
 
-    def shrink_quad_along_width(self, quad, begin_width_ratio=0., end_width_ratio=1.):
+    def shrink_quad_along_width(self,
+                                quad,
+                                begin_width_ratio=0.,
+                                end_width_ratio=1.):
         """ 
         Generate shrink_quad_along_width.
         """
-        ratio_pair = np.array([[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
+        ratio_pair = np.array(
+            [[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
         p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair
         p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair
         return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]])
@@ -81,16 +86,23 @@ class SASTPostProcess(object):
         expand poly along width.
         """
         point_num = poly.shape[0]
-        left_quad = np.array([poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32)
+        left_quad = np.array(
+            [poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32)
         left_ratio = -shrink_ratio_of_width * np.linalg.norm(left_quad[0] - left_quad[3]) / \
                      (np.linalg.norm(left_quad[0] - left_quad[1]) + 1e-6)
-        left_quad_expand = self.shrink_quad_along_width(left_quad, left_ratio, 1.0)
-        right_quad = np.array([poly[point_num // 2 - 2], poly[point_num // 2 - 1],
-                            poly[point_num // 2], poly[point_num // 2 + 1]], dtype=np.float32)
+        left_quad_expand = self.shrink_quad_along_width(left_quad, left_ratio,
+                                                        1.0)
+        right_quad = np.array(
+            [
+                poly[point_num // 2 - 2], poly[point_num // 2 - 1],
+                poly[point_num // 2], poly[point_num // 2 + 1]
+            ],
+            dtype=np.float32)
         right_ratio = 1.0 + \
                       shrink_ratio_of_width * np.linalg.norm(right_quad[0] - right_quad[3]) / \
                       (np.linalg.norm(right_quad[0] - right_quad[1]) + 1e-6)
-        right_quad_expand = self.shrink_quad_along_width(right_quad, 0.0, right_ratio)
+        right_quad_expand = self.shrink_quad_along_width(right_quad, 0.0,
+                                                         right_ratio)
         poly[0] = left_quad_expand[0]
         poly[-1] = left_quad_expand[-1]
         poly[point_num // 2 - 1] = right_quad_expand[1]
@@ -121,12 +133,10 @@ class SASTPostProcess(object):
         """
         compute area of a quad.
         """
-        edge = [
-            (quad[1][0] - quad[0][0]) * (quad[1][1] + quad[0][1]),
+        edge = [(quad[1][0] - quad[0][0]) * (quad[1][1] + quad[0][1]),
                 (quad[2][0] - quad[1][0]) * (quad[2][1] + quad[1][1]),
                 (quad[3][0] - quad[2][0]) * (quad[3][1] + quad[2][1]),
-            (quad[0][0] - quad[3][0]) * (quad[0][1] + quad[3][1])
-        ]
+                (quad[0][0] - quad[3][0]) * (quad[0][1] + quad[3][1])]
         return np.sum(edge) / 2.
 
     def nms(self, dets):
@@ -157,7 +167,8 @@ class SASTPostProcess(object):
         m = quads.shape[0]
         gt_tc = np.mean(quads, axis=1)  # (m, 2)
 
-        pred_tc_tile = np.tile(pred_tc[:, np.newaxis, :], (1, m, 1)) # (n, m, 2)
+        pred_tc_tile = np.tile(pred_tc[:, np.newaxis, :],
+                               (1, m, 1))  # (n, m, 2)
         gt_tc_tile = np.tile(gt_tc[np.newaxis, :, :], (n, 1, 1))  # (n, m, 2)
         dist_mat = np.linalg.norm(pred_tc_tile - gt_tc_tile, axis=2)  # (n, m)
         xy_text_assign = np.argmin(dist_mat, axis=1) + 1  # (n,)
@@ -169,26 +180,47 @@ class SASTPostProcess(object):
         """
         Estimate sample points number.
         """
-        eh = (np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[1] - quad[2])) / 2.0
-        ew = (np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[2] - quad[3])) / 2.0
+        eh = (np.linalg.norm(quad[0] - quad[3]) +
+              np.linalg.norm(quad[1] - quad[2])) / 2.0
+        ew = (np.linalg.norm(quad[0] - quad[1]) +
+              np.linalg.norm(quad[2] - quad[3])) / 2.0
 
         dense_sample_pts_num = max(2, int(ew))
-        dense_xy_center_line = xy_text[np.linspace(0, xy_text.shape[0] - 1, dense_sample_pts_num,
-                                                endpoint=True, dtype=np.float32).astype(np.int32)]
-
-        dense_xy_center_line_diff = dense_xy_center_line[1:] - dense_xy_center_line[:-1]
-        estimate_arc_len = np.sum(np.linalg.norm(dense_xy_center_line_diff, axis=1))
+        dense_xy_center_line = xy_text[np.linspace(
+            0,
+            xy_text.shape[0] - 1,
+            dense_sample_pts_num,
+            endpoint=True,
+            dtype=np.float32).astype(np.int32)]
+
+        dense_xy_center_line_diff = dense_xy_center_line[
+            1:] - dense_xy_center_line[:-1]
+        estimate_arc_len = np.sum(
+            np.linalg.norm(
+                dense_xy_center_line_diff, axis=1))
 
         sample_pts_num = max(2, int(estimate_arc_len / eh))
         return sample_pts_num
 
-    def detect_sast(self, tcl_map, tvo_map, tbo_map, tco_map, ratio_w, ratio_h, src_w, src_h, 
-                shrink_ratio_of_width=0.3, tcl_map_thresh=0.5, offset_expand=1.0, out_strid=4.0):
+    def detect_sast(self,
+                    tcl_map,
+                    tvo_map,
+                    tbo_map,
+                    tco_map,
+                    ratio_w,
+                    ratio_h,
+                    src_w,
+                    src_h,
+                    shrink_ratio_of_width=0.3,
+                    tcl_map_thresh=0.5,
+                    offset_expand=1.0,
+                    out_strid=4.0):
         """
         first resize the tcl_map, tvo_map and tbo_map to the input_size, then restore the polys
         """
         # restore quad
-        scores, quads, xy_text = self.restore_quad(tcl_map, tcl_map_thresh, tvo_map)
+        scores, quads, xy_text = self.restore_quad(tcl_map, tcl_map_thresh,
+                                                   tvo_map)
         dets = np.hstack((quads, scores)).astype(np.float32, copy=False)
         dets = self.nms(dets)
         if dets.shape[0] == 0:
@@ -202,7 +234,8 @@ class SASTPostProcess(object):
 
         # instance segmentation
         # instance_count, instance_label_map = cv2.connectedComponents(tcl_map.astype(np.uint8), connectivity=8)
-        instance_count, instance_label_map = self.cluster_by_quads_tco(tcl_map, tcl_map_thresh, quads, tco_map)
+        instance_count, instance_label_map = self.cluster_by_quads_tco(
+            tcl_map, tcl_map_thresh, quads, tco_map)
 
         # restore single poly with tcl instance.
         poly_list = []
@@ -214,8 +247,8 @@ class SASTPostProcess(object):
                 continue
 
             #
-            len1 = float(np.linalg.norm(quad[0] -quad[1]))
-            len2 = float(np.linalg.norm(quad[1] -quad[2]))
+            len1 = float(np.linalg.norm(quad[0] - quad[1]))
+            len2 = float(np.linalg.norm(quad[1] - quad[2]))
             min_len = min(len1, len2)
             if min_len < 3:
                 continue
@@ -231,9 +264,11 @@ class SASTPostProcess(object):
                 continue
 
             # sort xy_text
-            left_center_pt = np.array([[(quad[0, 0] + quad[-1, 0]) / 2.0,
+            left_center_pt = np.array(
+                [[(quad[0, 0] + quad[-1, 0]) / 2.0,
                   (quad[0, 1] + quad[-1, 1]) / 2.0]])  # (1, 2)
-            right_center_pt = np.array([[(quad[1, 0] + quad[2, 0]) / 2.0,
+            right_center_pt = np.array(
+                [[(quad[1, 0] + quad[2, 0]) / 2.0,
                   (quad[1, 1] + quad[2, 1]) / 2.0]])  # (1, 2)
             proj_unit_vec = (right_center_pt - left_center_pt) / \
                             (np.linalg.norm(right_center_pt - left_center_pt) + 1e-6)
@@ -245,28 +280,40 @@ class SASTPostProcess(object):
                 sample_pts_num = self.estimate_sample_pts_num(quad, xy_text)
             else:
                 sample_pts_num = self.sample_pts_num
-            xy_center_line = xy_text[np.linspace(0, xy_text.shape[0] - 1, sample_pts_num,
-                                                endpoint=True, dtype=np.float32).astype(np.int32)]
+            xy_center_line = xy_text[np.linspace(
+                0,
+                xy_text.shape[0] - 1,
+                sample_pts_num,
+                endpoint=True,
+                dtype=np.float32).astype(np.int32)]
 
             point_pair_list = []
             for x, y in xy_center_line:
                 # get corresponding offset
                 offset = tbo_map[y, x, :].reshape(2, 2)
                 if offset_expand != 1.0:
-                    offset_length = np.linalg.norm(offset, axis=1, keepdims=True)
-                    expand_length = np.clip(offset_length * (offset_expand - 1), a_min=0.5, a_max=3.0)
+                    offset_length = np.linalg.norm(
+                        offset, axis=1, keepdims=True)
+                    expand_length = np.clip(
+                        offset_length * (offset_expand - 1),
+                        a_min=0.5,
+                        a_max=3.0)
                     offset_detal = offset / offset_length * expand_length
                     offset = offset + offset_detal
                     # original point
                 ori_yx = np.array([y, x], dtype=np.float32)
-                point_pair = (ori_yx +  offset)[:, ::-1]* out_strid / np.array([ratio_w, ratio_h]).reshape(-1, 2) 
+                point_pair = (ori_yx + offset)[:, ::-1] * out_strid / np.array(
+                    [ratio_w, ratio_h]).reshape(-1, 2)
                 point_pair_list.append(point_pair)
 
             # ndarry: (x, 2), expand poly along width
             detected_poly = self.point_pair2poly(point_pair_list)
-            detected_poly = self.expand_poly_along_width(detected_poly, shrink_ratio_of_width)
-            detected_poly[:, 0] = np.clip(detected_poly[:, 0], a_min=0, a_max=src_w)
-            detected_poly[:, 1] = np.clip(detected_poly[:, 1], a_min=0, a_max=src_h)
+            detected_poly = self.expand_poly_along_width(detected_poly,
+                                                         shrink_ratio_of_width)
+            detected_poly[:, 0] = np.clip(
+                detected_poly[:, 0], a_min=0, a_max=src_w)
+            detected_poly[:, 1] = np.clip(
+                detected_poly[:, 1], a_min=0, a_max=src_h)
             poly_list.append(detected_poly)
 
         return poly_list
@@ -285,16 +332,24 @@ class SASTPostProcess(object):
         img_num = len(shape_list)
         poly_lists = []
         for ino in range(img_num):
-            p_score = score_list[ino].transpose((1,2,0))
-            p_border = border_list[ino].transpose((1,2,0))
-            p_tvo = tvo_list[ino].transpose((1,2,0))
-            p_tco = tco_list[ino].transpose((1,2,0))
+            p_score = score_list[ino].transpose((1, 2, 0))
+            p_border = border_list[ino].transpose((1, 2, 0))
+            p_tvo = tvo_list[ino].transpose((1, 2, 0))
+            p_tco = tco_list[ino].transpose((1, 2, 0))
             src_h, src_w, ratio_h, ratio_w = shape_list[ino]
 
-            poly_list = self.detect_sast(p_score, p_tvo, p_border, p_tco, ratio_w, ratio_h, src_w, src_h, 
+            poly_list = self.detect_sast(
+                p_score,
+                p_tvo,
+                p_border,
+                p_tco,
+                ratio_w,
+                ratio_h,
+                src_w,
+                src_h,
                 shrink_ratio_of_width=self.shrink_ratio_of_width,
-                                         tcl_map_thresh=self.tcl_map_thresh, offset_expand=self.expand_scale)
+                tcl_map_thresh=self.tcl_map_thresh,
+                offset_expand=self.expand_scale)
             poly_lists.append({'points': np.array(poly_list)})
 
         return poly_lists
-

ppocr/utils/dict/arabic_dict.txt

+ 
+!
+#
+$
+%
+&
+'
+(
++
+,
+-
+.
+/
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+:
+?
+@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
+_
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+É
+é
+ء
+آ
+أ
+ؤ
+إ
+ئ
+ا
+ب
+ة
+ت
+ث
+ج
+ح
+خ
+د
+ذ
+ر
+ز
+س
+ش
+ص
+ض
+ط
+ظ
+ع
+غ
+ف
+ق
+ك
+ل
+م
+ن
+ه
+و
+ى
+ي
+ً
+ٌ
+ٍ
+َ
+ُ
+ِ
+ّ
+ْ
+ٓ
+ٔ
+ٰ
+ٱ
+ٹ
+پ
+چ
+ڈ
+ڑ
+ژ
+ک
+ڭ
+گ
+ں
+ھ
+ۀ
+ہ
+ۂ
+ۃ
+ۆ
+ۇ
+ۈ
+ۋ
+ی
+ې
+ے
+ۓ
+ە
+١
+٢
+٣
+٤
+٥
+٦
+٧
+٨
+٩

ppocr/utils/dict/cyrillic_dict.txt

+ 
+!
+#
+$
+%
+&
+'
+(
++
+,
+-
+.
+/
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+:
+?
+@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
+_
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+É
+é
+Ё
+Є
+І
+Ј
+Љ
+Ў
+А
+Б
+В
+Г
+Д
+Е
+Ж
+З
+И
+Й
+К
+Л
+М
+Н
+О
+П
+Р
+С
+Т
+У
+Ф
+Х
+Ц
+Ч
+Ш
+Щ
+Ъ
+Ы
+Ь
+Э
+Ю
+Я
+а
+б
+в
+г
+д
+е
+ж
+з
+и
+й
+к
+л
+м
+н
+о
+п
+р
+с
+т
+у
+ф
+х
+ц
+ч
+ш
+щ
+ъ
+ы
+ь
+э
+ю
+я
+ё
+ђ
+є
+і
+ј
+љ
+њ
+ћ
+ў
+џ
+Ґ
+ґ

ppocr/utils/dict/devanagari_dict.txt

+ 
+!
+#
+$
+%
+&
+'
+(
++
+,
+-
+.
+/
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+:
+?
+@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
+_
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+É
+é
+ँ
+ं
+ः
+अ
+आ
+इ
+ई
+उ
+ऊ
+ऋ
+ए
+ऐ
+ऑ
+ओ
+औ
+क
+ख
+ग
+घ
+ङ
+च
+छ
+ज
+झ
+ञ
+ट
+ठ
+ड
+ढ
+ण
+त
+थ
+द
+ध
+न
+ऩ
+प
+फ
+ब
+भ
+म
+य
+र
+ऱ
+ल
+ळ
+व
+श
+ष
+स
+ह
+़
+ा
+ि
+ी
+ु
+ू
+ृ
+ॅ
+े
+ै
+ॉ
+ो
+ौ
+्
+॒
+क़
+ख़
+ग़
+ज़
+ड़
+ढ़
+फ़
+ॠ
+।
+०
+१
+२
+३
+४
+५
+६
+७
+८
+९
+॰

ppocr/utils/dict/latin_dict.txt

+ 
+!
+"
+#
+$
+%
+&
+'
+(
+)
+*
++
+,
+-
+.
+/
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+:
+;
+<
+=
+>
+?
+@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
+[
+]
+_
+`
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+{
+}
+¡
+£
+§
+ª
+«
+­
+°
+²
+³
+´
+µ
+·
+º
+»
+¿
+À
+Á
+Â
+Ä
+Å
+Ç
+È
+É
+Ê
+Ë
+Ì
+Í
+Î
+Ï
+Ò
+Ó
+Ô
+Õ
+Ö
+Ú
+Ü
+Ý
+ß
+à
+á
+â
+ã
+ä
+å
+æ
+ç
+è
+é
+ê
+ë
+ì
+í
+î
+ï
+ñ
+ò
+ó
+ô
+õ
+ö
+ø
+ù
+ú
+û
+ü
+ý
+ą
+Ć
+ć
+Č
+č
+Đ
+đ
+ę
+ı
+Ł
+ł
+ō
+Œ
+œ
+Š
+š
+Ÿ
+Ž
+ž
+ʒ
+β
+δ
+ε
+з
+Ṡ
+‘
+€
+™

ppocr/utils/e2e_metric/polygon_fast.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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
+#
+#     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
+from shapely.geometry import Polygon
+"""
+:param det_x: [1, N] Xs of detection's vertices 
+:param det_y: [1, N] Ys of detection's vertices
+:param gt_x: [1, N] Xs of groundtruth's vertices
+:param gt_y: [1, N] Ys of groundtruth's vertices
+
+##############
+All the calculation of 'AREA' in this script is handled by:
+1) First generating a binary mask with the polygon area filled up with 1's
+2) Summing up all the 1's
+"""
+
+
+def area(x, y):
+    polygon = Polygon(np.stack([x, y], axis=1))
+    return float(polygon.area)
+
+
+def approx_area_of_intersection(det_x, det_y, gt_x, gt_y):
+    """
+    This helper determine if both polygons are intersecting with each others with an approximation method.
+    Area of intersection represented by the minimum bounding rectangular [xmin, ymin, xmax, ymax]
+    """
+    det_ymax = np.max(det_y)
+    det_xmax = np.max(det_x)
+    det_ymin = np.min(det_y)
+    det_xmin = np.min(det_x)
+
+    gt_ymax = np.max(gt_y)
+    gt_xmax = np.max(gt_x)
+    gt_ymin = np.min(gt_y)
+    gt_xmin = np.min(gt_x)
+
+    all_min_ymax = np.minimum(det_ymax, gt_ymax)
+    all_max_ymin = np.maximum(det_ymin, gt_ymin)
+
+    intersect_heights = np.maximum(0.0, (all_min_ymax - all_max_ymin))
+
+    all_min_xmax = np.minimum(det_xmax, gt_xmax)
+    all_max_xmin = np.maximum(det_xmin, gt_xmin)
+    intersect_widths = np.maximum(0.0, (all_min_xmax - all_max_xmin))
+
+    return intersect_heights * intersect_widths
+
+
+def area_of_intersection(det_x, det_y, gt_x, gt_y):
+    p1 = Polygon(np.stack([det_x, det_y], axis=1)).buffer(0)
+    p2 = Polygon(np.stack([gt_x, gt_y], axis=1)).buffer(0)
+    return float(p1.intersection(p2).area)
+
+
+def area_of_union(det_x, det_y, gt_x, gt_y):
+    p1 = Polygon(np.stack([det_x, det_y], axis=1)).buffer(0)
+    p2 = Polygon(np.stack([gt_x, gt_y], axis=1)).buffer(0)
+    return float(p1.union(p2).area)
+
+
+def iou(det_x, det_y, gt_x, gt_y):
+    return area_of_intersection(det_x, det_y, gt_x, gt_y) / (
+        area_of_union(det_x, det_y, gt_x, gt_y) + 1.0)
+
+
+def iod(det_x, det_y, gt_x, gt_y):
+    """
+    This helper determine the fraction of intersection area over detection area
+    """
+    return area_of_intersection(det_x, det_y, gt_x, gt_y) / (
+        area(det_x, det_y) + 1.0)

ppocr/utils/e2e_utils/extract_batchsize.py

+import paddle
+import numpy as np
+import copy
+
+
+def org_tcl_rois(batch_size, pos_lists, pos_masks, label_lists, tcl_bs):
+    """
+    """
+    pos_lists_, pos_masks_, label_lists_ = [], [], []
+    img_bs = batch_size
+    ngpu = int(batch_size / img_bs)
+    img_ids = np.array(pos_lists, dtype=np.int32)[:, 0, 0].copy()
+    pos_lists_split, pos_masks_split, label_lists_split = [], [], []
+    for i in range(ngpu):
+        pos_lists_split.append([])
+        pos_masks_split.append([])
+        label_lists_split.append([])
+
+    for i in range(img_ids.shape[0]):
+        img_id = img_ids[i]
+        gpu_id = int(img_id / img_bs)
+        img_id = img_id % img_bs
+        pos_list = pos_lists[i].copy()
+        pos_list[:, 0] = img_id
+        pos_lists_split[gpu_id].append(pos_list)
+        pos_masks_split[gpu_id].append(pos_masks[i].copy())
+        label_lists_split[gpu_id].append(copy.deepcopy(label_lists[i]))
+    # repeat or delete
+    for i in range(ngpu):
+        vp_len = len(pos_lists_split[i])
+        if vp_len <= tcl_bs:
+            for j in range(0, tcl_bs - vp_len):
+                pos_list = pos_lists_split[i][j].copy()
+                pos_lists_split[i].append(pos_list)
+                pos_mask = pos_masks_split[i][j].copy()
+                pos_masks_split[i].append(pos_mask)
+                label_list = copy.deepcopy(label_lists_split[i][j])
+                label_lists_split[i].append(label_list)
+        else:
+            for j in range(0, vp_len - tcl_bs):
+                c_len = len(pos_lists_split[i])
+                pop_id = np.random.permutation(c_len)[0]
+                pos_lists_split[i].pop(pop_id)
+                pos_masks_split[i].pop(pop_id)
+                label_lists_split[i].pop(pop_id)
+    # merge
+    for i in range(ngpu):
+        pos_lists_.extend(pos_lists_split[i])
+        pos_masks_.extend(pos_masks_split[i])
+        label_lists_.extend(label_lists_split[i])
+    return pos_lists_, pos_masks_, label_lists_
+
+
+def pre_process(label_list, pos_list, pos_mask, max_text_length, max_text_nums,
+                pad_num, tcl_bs):
+    label_list = label_list.numpy()
+    batch, _, _, _ = label_list.shape
+    pos_list = pos_list.numpy()
+    pos_mask = pos_mask.numpy()
+    pos_list_t = []
+    pos_mask_t = []
+    label_list_t = []
+    for i in range(batch):
+        for j in range(max_text_nums):
+            if pos_mask[i, j].any():
+                pos_list_t.append(pos_list[i][j])
+                pos_mask_t.append(pos_mask[i][j])
+                label_list_t.append(label_list[i][j])
+    pos_list, pos_mask, label_list = org_tcl_rois(batch, pos_list_t, pos_mask_t,
+                                                  label_list_t, tcl_bs)
+    label = []
+    tt = [l.tolist() for l in label_list]
+    for i in range(tcl_bs):
+        k = 0
+        for j in range(max_text_length):
+            if tt[i][j][0] != pad_num:
+                k += 1
+            else:
+                break
+        label.append(k)
+    label = paddle.to_tensor(label)
+    label = paddle.cast(label, dtype='int64')
+    pos_list = paddle.to_tensor(pos_list)
+    pos_mask = paddle.to_tensor(pos_mask)
+    label_list = paddle.squeeze(paddle.to_tensor(label_list), axis=2)
+    label_list = paddle.cast(label_list, dtype='int32')
+    return pos_list, pos_mask, label_list, label

ppocr/utils/e2e_utils/visual.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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
+#
+#     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 cv2
+import time
+
+
+def resize_image(im, max_side_len=512):
+    """
+    resize image to a size multiple of max_stride which is required by the network
+    :param im: the resized image
+    :param max_side_len: limit of max image size to avoid out of memory in gpu
+    :return: the resized image and the resize ratio
+    """
+    h, w, _ = im.shape
+
+    resize_w = w
+    resize_h = h
+
+    if resize_h > resize_w:
+        ratio = float(max_side_len) / resize_h
+    else:
+        ratio = float(max_side_len) / resize_w
+
+    resize_h = int(resize_h * ratio)
+    resize_w = int(resize_w * ratio)
+
+    max_stride = 128
+    resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+    resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+    im = cv2.resize(im, (int(resize_w), int(resize_h)))
+    ratio_h = resize_h / float(h)
+    ratio_w = resize_w / float(w)
+
+    return im, (ratio_h, ratio_w)
+
+
+def resize_image_min(im, max_side_len=512):
+    """
+    """
+    h, w, _ = im.shape
+
+    resize_w = w
+    resize_h = h
+
+    if resize_h < resize_w:
+        ratio = float(max_side_len) / resize_h
+    else:
+        ratio = float(max_side_len) / resize_w
+
+    resize_h = int(resize_h * ratio)
+    resize_w = int(resize_w * ratio)
+
+    max_stride = 128
+    resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+    resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+    im = cv2.resize(im, (int(resize_w), int(resize_h)))
+    ratio_h = resize_h / float(h)
+    ratio_w = resize_w / float(w)
+    return im, (ratio_h, ratio_w)
+
+
+def resize_image_for_totaltext(im, max_side_len=512):
+    """
+    """
+    h, w, _ = im.shape
+
+    resize_w = w
+    resize_h = h
+    ratio = 1.25
+    if h * ratio > max_side_len:
+        ratio = float(max_side_len) / resize_h
+
+    resize_h = int(resize_h * ratio)
+    resize_w = int(resize_w * ratio)
+
+    max_stride = 128
+    resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+    resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+    im = cv2.resize(im, (int(resize_w), int(resize_h)))
+    ratio_h = resize_h / float(h)
+    ratio_w = resize_w / float(w)
+    return im, (ratio_h, ratio_w)
+
+
+def point_pair2poly(point_pair_list):
+    """
+    Transfer vertical point_pairs into poly point in clockwise.
+    """
+    pair_length_list = []
+    for point_pair in point_pair_list:
+        pair_length = np.linalg.norm(point_pair[0] - point_pair[1])
+        pair_length_list.append(pair_length)
+    pair_length_list = np.array(pair_length_list)
+    pair_info = (pair_length_list.max(), pair_length_list.min(),
+                 pair_length_list.mean())
+
+    point_num = len(point_pair_list) * 2
+    point_list = [0] * point_num
+    for idx, point_pair in enumerate(point_pair_list):
+        point_list[idx] = point_pair[0]
+        point_list[point_num - 1 - idx] = point_pair[1]
+    return np.array(point_list).reshape(-1, 2), pair_info
+
+
+def shrink_quad_along_width(quad, begin_width_ratio=0., end_width_ratio=1.):
+    """
+    Generate shrink_quad_along_width.
+    """
+    ratio_pair = np.array(
+        [[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
+    p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair
+    p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair
+    return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]])
+
+
+def expand_poly_along_width(poly, shrink_ratio_of_width=0.3):
+    """
+    expand poly along width.
+    """
+    point_num = poly.shape[0]
+    left_quad = np.array(
+        [poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32)
+    left_ratio = -shrink_ratio_of_width * np.linalg.norm(left_quad[0] - left_quad[3]) / \
+                 (np.linalg.norm(left_quad[0] - left_quad[1]) + 1e-6)
+    left_quad_expand = shrink_quad_along_width(left_quad, left_ratio, 1.0)
+    right_quad = np.array(
+        [
+            poly[point_num // 2 - 2], poly[point_num // 2 - 1],
+            poly[point_num // 2], poly[point_num // 2 + 1]
+        ],
+        dtype=np.float32)
+    right_ratio = 1.0 + \
+                  shrink_ratio_of_width * np.linalg.norm(right_quad[0] - right_quad[3]) / \
+                  (np.linalg.norm(right_quad[0] - right_quad[1]) + 1e-6)
+    right_quad_expand = shrink_quad_along_width(right_quad, 0.0, right_ratio)
+    poly[0] = left_quad_expand[0]
+    poly[-1] = left_quad_expand[-1]
+    poly[point_num // 2 - 1] = right_quad_expand[1]
+    poly[point_num // 2] = right_quad_expand[2]
+    return poly
+
+
+def norm2(x, axis=None):
+    if axis:
+        return np.sqrt(np.sum(x**2, axis=axis))
+    return np.sqrt(np.sum(x**2))
+
+
+def cos(p1, p2):
+    return (p1 * p2).sum() / (norm2(p1) * norm2(p2))

requirements.txt

@@ -8,3 +8,4 @@ tqdm
 numpy
 visualdl
 python-Levenshtein
+opencv-contrib-python
\ No newline at end of file

setup.py

@@ -32,7 +32,7 @@ setup(
     package_dir={'paddleocr': ''},
     include_package_data=True,
     entry_points={"console_scripts": ["paddleocr= paddleocr.paddleocr:main"]},
-    version='2.0.3',
+    version='2.0.4',
     install_requires=requirements,
     license='Apache License 2.0',
     description='Awesome OCR toolkits based on PaddlePaddle （8.6M ultra-lightweight pre-trained model, support training and deployment among server, mobile, embeded and IoT devices',

tools/infer/utility.py

@@ -74,6 +74,19 @@ def parse_args():
         "--vis_font_path", type=str, default="./doc/fonts/simfang.ttf")
     parser.add_argument("--drop_score", type=float, default=0.5)
 
+    # params for e2e
+    parser.add_argument("--e2e_algorithm", type=str, default='PGNet')
+    parser.add_argument("--e2e_model_dir", type=str)
+    parser.add_argument("--e2e_limit_side_len", type=float, default=768)
+    parser.add_argument("--e2e_limit_type", type=str, default='max')
+
+    # PGNet parmas
+    parser.add_argument("--e2e_pgnet_score_thresh", type=float, default=0.5)
+    parser.add_argument(
+        "--e2e_char_dict_path", type=str, default="./ppocr/utils/ic15_dict.txt")
+    parser.add_argument("--e2e_pgnet_valid_set", type=str, default='totaltext')
+    parser.add_argument("--e2e_pgnet_polygon", type=bool, default=True)
+
     # params for text classifier
     parser.add_argument("--use_angle_cls", type=str2bool, default=False)
     parser.add_argument("--cls_model_dir", type=str)
@@ -93,8 +106,10 @@ def create_predictor(args, mode, logger):
         model_dir = args.det_model_dir
     elif mode == 'cls':
         model_dir = args.cls_model_dir
-    else:
+    elif mode == 'rec':
         model_dir = args.rec_model_dir
+    else:
+        model_dir = args.e2e_model_dir
 
     if model_dir is None:
         logger.info("not find {} model file path {}".format(mode, model_dir))
@@ -148,6 +163,22 @@ def create_predictor(args, mode, logger):
     return predictor, input_tensor, output_tensors
 
 
+def draw_e2e_res(dt_boxes, strs, img_path):
+    src_im = cv2.imread(img_path)
+    for box, str in zip(dt_boxes, strs):
+        box = box.astype(np.int32).reshape((-1, 1, 2))
+        cv2.polylines(src_im, [box], True, color=(255, 255, 0), thickness=2)
+        cv2.putText(
+            src_im,
+            str,
+            org=(int(box[0, 0, 0]), int(box[0, 0, 1])),
+            fontFace=cv2.FONT_HERSHEY_COMPLEX,
+            fontScale=0.7,
+            color=(0, 255, 0),
+            thickness=1)
+    return src_im
+
+
 def draw_text_det_res(dt_boxes, img_path):
     src_im = cv2.imread(img_path)
     for box in dt_boxes: