m53297601
/
PaddleOCR
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #1363 from MissPenguin/dygraph

Browse Source 
add east & sast
release/2.0-rc1-0
MissPenguin 4 years ago committed by GitHub
parent 3c9d3f6bf5 d42bf7a0b7
commit bd7f8f72cc
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
26 changed files with 3468 additions and 13 deletions
Show Stats Download Patch File Download Diff File

111
configs/det/det_mv3_east.yml
Unescape View File

@ -0,0 +1,111 @@
Global:
use_gpu: true
epoch_num: 10000
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/east_mv3/
save_epoch_step: 1000
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
# if pretrained_model is saved in static mode, load_static_weights must set to True
load_static_weights: True
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img:
save_res_path: ./output/det_east/predicts_east.txt
Architecture:
model_type: det
algorithm: EAST
Transform:
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
Neck:
name: EASTFPN
model_name: small
Head:
name: EASTHead
model_name: small
Loss:
name: EASTLoss
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
# name: Cosine
learning_rate: 0.001
# warmup_epoch: 0
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: EASTPostProcess
score_thresh: 0.8
cover_thresh: 0.1
nms_thresh: 0.2
Metric:
name: DetMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
ratio_list: [1.0]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- EASTProcessTrain:
image_shape: [512, 512]
background_ratio: 0.125
min_crop_side_ratio: 0.1
min_text_size: 10
- KeepKeys:
keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 16
num_workers: 8
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
limit_side_len: 2400
limit_type: max
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2

110
configs/det/det_r50_vd_east.yml
Unescape View File

@ -0,0 +1,110 @@
Global:
use_gpu: true
epoch_num: 10000
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/east_r50_vd/
save_epoch_step: 1000
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
# if pretrained_model is saved in static mode, load_static_weights must set to True
load_static_weights: True
cal_metric_during_train: False
pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img:
save_res_path: ./output/det_east/predicts_east.txt
Architecture:
model_type: det
algorithm: EAST
Transform:
Backbone:
name: ResNet
layers: 50
Neck:
name: EASTFPN
model_name: large
Head:
name: EASTHead
model_name: large
Loss:
name: EASTLoss
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
# name: Cosine
learning_rate: 0.001
# warmup_epoch: 0
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: EASTPostProcess
score_thresh: 0.8
cover_thresh: 0.1
nms_thresh: 0.2
Metric:
name: DetMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
ratio_list: [1.0]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- EASTProcessTrain:
image_shape: [512, 512]
background_ratio: 0.125
min_crop_side_ratio: 0.1
min_text_size: 10
- KeepKeys:
keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 8
num_workers: 8
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
limit_side_len: 2400
limit_type: max
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2

110
configs/det/det_r50_vd_sast_icdar15.yml
Unescape View File

@ -0,0 +1,110 @@
Global:
use_gpu: true
epoch_num: 5000
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/sast_r50_vd_ic15/
save_epoch_step: 1000
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
# if pretrained_model is saved in static mode, load_static_weights must set to True
load_static_weights: True
cal_metric_during_train: False
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img:
save_res_path: ./output/sast_r50_vd_ic15/predicts_sast.txt
Architecture:
model_type: det
algorithm: SAST
Transform:
Backbone:
name: ResNet_SAST
layers: 50
Neck:
name: SASTFPN
with_cab: True
Head:
name: SASTHead
Loss:
name: SASTLoss
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
# name: Cosine
learning_rate: 0.001
# warmup_epoch: 0
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: SASTPostProcess
score_thresh: 0.5
sample_pts_num: 2
nms_thresh: 0.2
expand_scale: 1.0
shrink_ratio_of_width: 0.3
Metric:
name: DetMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
label_file_path: [./train_data/art_latin_icdar_14pt/train_no_tt_test/train_label_json.txt, ./train_data/total_text_icdar_14pt/train_label_json.txt]
data_ratio_list: [0.5, 0.5]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- SASTProcessTrain:
image_shape: [512, 512]
min_crop_side_ratio: 0.3
min_crop_size: 24
min_text_size: 4
max_text_size: 512
- KeepKeys:
keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 4
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
resize_long: 1536
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2

109
configs/det/det_r50_vd_sast_totaltext.yml
Unescape View File

@ -0,0 +1,109 @@
Global:
use_gpu: true
epoch_num: 5000
log_smooth_window: 20
print_batch_step: 2
save_model_dir: ./output/sast_r50_vd_tt/
save_epoch_step: 1000
# evaluation is run every 5000 iterations after the 4000th iteration
eval_batch_step: [4000, 5000]
# if pretrained_model is saved in static mode, load_static_weights must set to True
load_static_weights: True
cal_metric_during_train: False
pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img:
save_res_path: ./output/sast_r50_vd_tt/predicts_sast.txt
Architecture:
model_type: det
algorithm: SAST
Transform:
Backbone:
name: ResNet_SAST
layers: 50
Neck:
name: SASTFPN
with_cab: True
Head:
name: SASTHead
Loss:
name: SASTLoss
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
# name: Cosine
learning_rate: 0.001
# warmup_epoch: 0
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: SASTPostProcess
score_thresh: 0.5
sample_pts_num: 6
nms_thresh: 0.2
expand_scale: 1.2
shrink_ratio_of_width: 0.2
Metric:
name: DetMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
label_file_list: [./train_data/icdar2013/train_label_json.txt, ./train_data/icdar2015/train_label_json.txt, ./train_data/icdar17_mlt_latin/train_label_json.txt, ./train_data/coco_text_icdar_4pts/train_label_json.txt]
ratio_list: [0.1, 0.45, 0.3, 0.15]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- SASTProcessTrain:
image_shape: [512, 512]
min_crop_side_ratio: 0.3
min_crop_size: 24
min_text_size: 4
max_text_size: 512
- KeepKeys:
keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order
loader:
shuffle: True
drop_last: False
batch_size_per_card: 4
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
label_file_list:
- ./train_data/total_text_icdar_14pt/test_label_json.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
resize_long: 768
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 2

3
ppocr/data/imaug/__init__.py
Unescape View File

@ -26,6 +26,9 @@ from .randaugment import RandAugment
from .operators import *
from .label_ops import *
from .east_process import *
from .sast_process import *
def transform(data, ops=None):
""" transform """

439
ppocr/data/imaug/east_process.py
View File

File diff suppressed because it is too large Load Diff

14
ppocr/data/imaug/label_ops.py
Unescape View File

@ -52,6 +52,7 @@ class DetLabelEncode(object):
txt_tags.append(True)
else:
txt_tags.append(False)
boxes = self.expand_points_num(boxes)
boxes = np.array(boxes, dtype=np.float32)
txt_tags = np.array(txt_tags, dtype=np.bool)
@ -70,6 +71,17 @@ class DetLabelEncode(object):
rect[3] = pts[np.argmax(diff)]
return rect
def expand_points_num(self, boxes):
max_points_num = 0
for box in boxes:
if len(box) > max_points_num:
max_points_num = len(box)
ex_boxes = []
for box in boxes:
ex_box = box + [box[-1]] * (max_points_num - len(box))
ex_boxes.append(ex_box)
return ex_boxes
class BaseRecLabelEncode(object):
""" Convert between text-label and text-index """
@ -83,7 +95,7 @@ class BaseRecLabelEncode(object):
'ch', 'en', 'en_sensitive', 'french', 'german', 'japan', 'korean'
]
assert character_type in support_character_type, "Only {} are supported now but get {}".format(
support_character_type, self.character_str)
support_character_type, character_type)
self.max_text_len = max_text_length
if character_type == "en":

48
ppocr/data/imaug/operators.py
Unescape View File

@ -120,26 +120,37 @@ class DetResizeForTest(object):
if 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min')
if 'resize_long' in kwargs:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
else:
self.limit_side_len = 736
self.limit_type = 'min'
def __call__(self, data):
img = data['image']
src_h, src_w, _ = img.shape
if self.resize_type == 0:
img, shape = self.resize_image_type0(img)
# img, shape = self.resize_image_type0(img)
img, [ratio_h, ratio_w] = self.resize_image_type0(img)
elif self.resize_type == 2:
img, [ratio_h, ratio_w] = self.resize_image_type2(img)
else:
img, shape = self.resize_image_type1(img)
# img, shape = self.resize_image_type1(img)
img, [ratio_h, ratio_w] = self.resize_image_type1(img)
data['image'] = img
data['shape'] = shape
data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
return data
def resize_image_type1(self, img):
resize_h, resize_w = self.image_shape
ori_h, ori_w = img.shape[:2] # (h, w, c)
ratio_h = float(resize_h) / ori_h
ratio_w = float(resize_w) / ori_w
img = cv2.resize(img, (int(resize_w), int(resize_h)))
return img, np.array([ori_h, ori_w])
# return img, np.array([ori_h, ori_w])
return img, [ratio_h, ratio_w]
def resize_image_type0(self, img):
"""
@ -182,4 +193,31 @@ class DetResizeForTest(object):
except:
print(img.shape, resize_w, resize_h)
sys.exit(0)
return img, np.array([h, w])
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
# return img, np.array([h, w])
return img, [ratio_h, ratio_w]
def resize_image_type2(self, img):
h, w, _ = img.shape
resize_w = w
resize_h = h
# Fix the longer side
if resize_h > resize_w:
ratio = float(self.resize_long) / resize_h
else:
ratio = float(self.resize_long) / 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
img = cv2.resize(img, (int(resize_w), int(resize_h)))
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
return img, [ratio_h, ratio_w]

689
ppocr/data/imaug/sast_process.py
View File

File diff suppressed because it is too large Load Diff

4
ppocr/losses/__init__.py
Unescape View File

@ -18,6 +18,8 @@ import copy
def build_loss(config):
# det loss
from .det_db_loss import DBLoss
from .det_east_loss import EASTLoss
from .det_sast_loss import SASTLoss
# rec loss
from .rec_ctc_loss import CTCLoss
@ -25,7 +27,7 @@ def build_loss(config):
# cls loss
from .cls_loss import ClsLoss
support_dict = ['DBLoss', 'CTCLoss', 'ClsLoss']
support_dict = ['DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss']
config = copy.deepcopy(config)
module_name = config.pop('name')

63
ppocr/losses/det_east_loss.py
Unescape View File

@ -0,0 +1,63 @@
# copyright (c) 2019 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
from .det_basic_loss import DiceLoss
class EASTLoss(nn.Layer):
"""
"""
def __init__(self,
eps=1e-6,
**kwargs):
super(EASTLoss, self).__init__()
self.dice_loss = DiceLoss(eps=eps)
def forward(self, predicts, labels):
l_score, l_geo, l_mask = labels[1:]
f_score = predicts['f_score']
f_geo = predicts['f_geo']
dice_loss = self.dice_loss(f_score, l_score, l_mask)
#smoooth_l1_loss
channels = 8
l_geo_split = paddle.split(
l_geo, num_or_sections=channels + 1, axis=1)
f_geo_split = paddle.split(f_geo, num_or_sections=channels, axis=1)
smooth_l1 = 0
for i in range(0, channels):
geo_diff = l_geo_split[i] - f_geo_split[i]
abs_geo_diff = paddle.abs(geo_diff)
smooth_l1_sign = paddle.less_than(abs_geo_diff, l_score)
smooth_l1_sign = paddle.cast(smooth_l1_sign, dtype='float32')
in_loss = abs_geo_diff * abs_geo_diff * smooth_l1_sign + \
(abs_geo_diff - 0.5) * (1.0 - smooth_l1_sign)
out_loss = l_geo_split[-1] / channels * in_loss * l_score
smooth_l1 += out_loss
smooth_l1_loss = paddle.mean(smooth_l1 * l_score)
dice_loss = dice_loss * 0.01
total_loss = dice_loss + smooth_l1_loss
losses = {"loss":total_loss, \
"dice_loss":dice_loss,\
"smooth_l1_loss":smooth_l1_loss}
return losses

121
ppocr/losses/det_sast_loss.py
Unescape View File

@ -0,0 +1,121 @@
# copyright (c) 2019 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
from .det_basic_loss import DiceLoss
import paddle.fluid as fluid
import numpy as np
class SASTLoss(nn.Layer):
"""
"""
def __init__(self,
eps=1e-6,
**kwargs):
super(SASTLoss, self).__init__()
self.dice_loss = DiceLoss(eps=eps)
def forward(self, predicts, labels):
"""
tcl_pos: N x 128 x 3
tcl_mask: N x 128 x 1
tcl_label: N x X list or LoDTensor
"""
f_score = predicts['f_score']
f_border = predicts['f_border']
f_tvo = predicts['f_tvo']
f_tco = predicts['f_tco']
l_score, l_border, l_mask, l_tvo, l_tco = labels[1:]
#score_loss
intersection = paddle.sum(f_score * l_score * l_mask)
union = paddle.sum(f_score * l_mask) + paddle.sum(l_score * l_mask)
score_loss = 1.0 - 2 * intersection / (union + 1e-5)
#border loss
l_border_split, l_border_norm = paddle.split(l_border, num_or_sections=[4, 1], axis=1)
f_border_split = f_border
border_ex_shape = l_border_norm.shape * np.array([1, 4, 1, 1])
l_border_norm_split = paddle.expand(x=l_border_norm, shape=border_ex_shape)
l_border_score = paddle.expand(x=l_score, shape=border_ex_shape)
l_border_mask = paddle.expand(x=l_mask, shape=border_ex_shape)
border_diff = l_border_split - f_border_split
abs_border_diff = paddle.abs(border_diff)
border_sign = abs_border_diff < 1.0
border_sign = paddle.cast(border_sign, dtype='float32')
border_sign.stop_gradient = True
border_in_loss = 0.5 * abs_border_diff * abs_border_diff * border_sign + \
(abs_border_diff - 0.5) * (1.0 - border_sign)
border_out_loss = l_border_norm_split * border_in_loss
border_loss = paddle.sum(border_out_loss * l_border_score * l_border_mask) / \
(paddle.sum(l_border_score * l_border_mask) + 1e-5)
#tvo_loss
l_tvo_split, l_tvo_norm = paddle.split(l_tvo, num_or_sections=[8, 1], axis=1)
f_tvo_split = f_tvo
tvo_ex_shape = l_tvo_norm.shape * np.array([1, 8, 1, 1])
l_tvo_norm_split = paddle.expand(x=l_tvo_norm, shape=tvo_ex_shape)
l_tvo_score = paddle.expand(x=l_score, shape=tvo_ex_shape)
l_tvo_mask = paddle.expand(x=l_mask, shape=tvo_ex_shape)
#
tvo_geo_diff = l_tvo_split - f_tvo_split
abs_tvo_geo_diff = paddle.abs(tvo_geo_diff)
tvo_sign = abs_tvo_geo_diff < 1.0
tvo_sign = paddle.cast(tvo_sign, dtype='float32')
tvo_sign.stop_gradient = True
tvo_in_loss = 0.5 * abs_tvo_geo_diff * abs_tvo_geo_diff * tvo_sign + \
(abs_tvo_geo_diff - 0.5) * (1.0 - tvo_sign)
tvo_out_loss = l_tvo_norm_split * tvo_in_loss
tvo_loss = paddle.sum(tvo_out_loss * l_tvo_score * l_tvo_mask) / \
(paddle.sum(l_tvo_score * l_tvo_mask) + 1e-5)
#tco_loss
l_tco_split, l_tco_norm = paddle.split(l_tco, num_or_sections=[2, 1], axis=1)
f_tco_split = f_tco
tco_ex_shape = l_tco_norm.shape * np.array([1, 2, 1, 1])
l_tco_norm_split = paddle.expand(x=l_tco_norm, shape=tco_ex_shape)
l_tco_score = paddle.expand(x=l_score, shape=tco_ex_shape)
l_tco_mask = paddle.expand(x=l_mask, shape=tco_ex_shape)
tco_geo_diff = l_tco_split - f_tco_split
abs_tco_geo_diff = paddle.abs(tco_geo_diff)
tco_sign = abs_tco_geo_diff < 1.0
tco_sign = paddle.cast(tco_sign, dtype='float32')
tco_sign.stop_gradient = True
tco_in_loss = 0.5 * abs_tco_geo_diff * abs_tco_geo_diff * tco_sign + \
(abs_tco_geo_diff - 0.5) * (1.0 - tco_sign)
tco_out_loss = l_tco_norm_split * tco_in_loss
tco_loss = paddle.sum(tco_out_loss * l_tco_score * l_tco_mask) / \
(paddle.sum(l_tco_score * l_tco_mask) + 1e-5)
# total loss
tvo_lw, tco_lw = 1.5, 1.5
score_lw, border_lw = 1.0, 1.0
total_loss = score_loss * score_lw + border_loss * border_lw + \
tvo_loss * tvo_lw + tco_loss * tco_lw
losses = {'loss':total_loss, "score_loss":score_loss,\
"border_loss":border_loss, 'tvo_loss':tvo_loss, 'tco_loss':tco_loss}
return losses

1
ppocr/modeling/backbones/__init__.py
Unescape View File

@ -19,6 +19,7 @@ def build_backbone(config, model_type):
if model_type == 'det':
from .det_mobilenet_v3 import MobileNetV3
from .det_resnet_vd import ResNet
from .det_resnet_vd_sast import ResNet_SAST
support_dict = ['MobileNetV3', 'ResNet', 'ResNet_SAST']
elif model_type == 'rec' or model_type == 'cls':
from .rec_mobilenet_v3 import MobileNetV3

285
ppocr/modeling/backbones/det_resnet_vd_sast.py
View File

File diff suppressed because it is too large Load Diff

4
ppocr/modeling/heads/__init__.py
Unescape View File

@ -18,13 +18,15 @@ __all__ = ['build_head']
def build_head(config):
# det head
from .det_db_head import DBHead
from .det_east_head import EASTHead
from .det_sast_head import SASTHead
# rec head
from .rec_ctc_head import CTCHead
# cls head
from .cls_head import ClsHead
support_dict = ['DBHead', 'CTCHead', 'ClsHead']
support_dict = ['DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead']
module_name = config.pop('name')
assert module_name in support_dict, Exception('head only support {}'.format(

121
ppocr/modeling/heads/det_east_head.py
Unescape View File

@ -0,0 +1,121 @@
# copyright (c) 2019 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")
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return x
class EASTHead(nn.Layer):
"""
"""
def __init__(self, in_channels, model_name, **kwargs):
super(EASTHead, self).__init__()
self.model_name = model_name
if self.model_name == "large":
num_outputs = [128, 64, 1, 8]
else:
num_outputs = [64, 32, 1, 8]
self.det_conv1 = ConvBNLayer(
in_channels=in_channels,
out_channels=num_outputs[0],
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="det_head1")
self.det_conv2 = ConvBNLayer(
in_channels=num_outputs[0],
out_channels=num_outputs[1],
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="det_head2")
self.score_conv = ConvBNLayer(
in_channels=num_outputs[1],
out_channels=num_outputs[2],
kernel_size=1,
stride=1,
padding=0,
if_act=False,
act=None,
name="f_score")
self.geo_conv = ConvBNLayer(
in_channels=num_outputs[1],
out_channels=num_outputs[3],
kernel_size=1,
stride=1,
padding=0,
if_act=False,
act=None,
name="f_geo")
def forward(self, x):
f_det = self.det_conv1(x)
f_det = self.det_conv2(f_det)
f_score = self.score_conv(f_det)
f_score = F.sigmoid(f_score)
f_geo = self.geo_conv(f_det)
f_geo = (F.sigmoid(f_geo) - 0.5) * 2 * 800
pred = {'f_score': f_score, 'f_geo': f_geo}
return pred

128
ppocr/modeling/heads/det_sast_head.py
Unescape View File

@ -0,0 +1,128 @@
# copyright (c) 2019 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,
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=(kernel_size - 1) // 2,
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")
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return x
class SAST_Header1(nn.Layer):
def __init__(self, in_channels, **kwargs):
super(SAST_Header1, self).__init__()
out_channels = [64, 64, 128]
self.score_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_score1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_score2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_score3'),
ConvBNLayer(out_channels[2], 1, 3, 1, act=None, name='f_score4')
)
self.border_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_border1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_border2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_border3'),
ConvBNLayer(out_channels[2], 4, 3, 1, act=None, name='f_border4')
)
def forward(self, x):
f_score = self.score_conv(x)
f_score = F.sigmoid(f_score)
f_border = self.border_conv(x)
return f_score, f_border
class SAST_Header2(nn.Layer):
def __init__(self, in_channels, **kwargs):
super(SAST_Header2, self).__init__()
out_channels = [64, 64, 128]
self.tvo_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tvo1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tvo2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tvo3'),
ConvBNLayer(out_channels[2], 8, 3, 1, act=None, name='f_tvo4')
)
self.tco_conv = nn.Sequential(
ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tco1'),
ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tco2'),
ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tco3'),
ConvBNLayer(out_channels[2], 2, 3, 1, act=None, name='f_tco4')
)
def forward(self, x):
f_tvo = self.tvo_conv(x)
f_tco = self.tco_conv(x)
return f_tvo, f_tco
class SASTHead(nn.Layer):
"""
"""
def __init__(self, in_channels, **kwargs):
super(SASTHead, self).__init__()
self.head1 = SAST_Header1(in_channels)
self.head2 = SAST_Header2(in_channels)
def forward(self, x):
f_score, f_border = self.head1(x)
f_tvo, f_tco = self.head2(x)
predicts = {}
predicts['f_score'] = f_score
predicts['f_border'] = f_border
predicts['f_tvo'] = f_tvo
predicts['f_tco'] = f_tco
return predicts

4
ppocr/modeling/necks/__init__.py
Unescape View File

@ -16,8 +16,10 @@ __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', 'SequenceEncoder']
support_dict = ['DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder']
module_name = config.pop('name')
assert module_name in support_dict, Exception('neck only support {}'.format(

188
ppocr/modeling/necks/east_fpn.py
Unescape View File

@ -0,0 +1,188 @@
# copyright (c) 2019 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,
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")
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return x
class DeConvBNLayer(nn.Layer):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
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")
def forward(self, x):
x = self.deconv(x)
x = self.bn(x)
return x
class EASTFPN(nn.Layer):
def __init__(self, in_channels, model_name, **kwargs):
super(EASTFPN, self).__init__()
self.model_name = model_name
if self.model_name == "large":
self.out_channels = 128
else:
self.out_channels = 64
self.in_channels = in_channels[::-1]
self.h1_conv = ConvBNLayer(
in_channels=self.out_channels+self.in_channels[1],
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="unet_h_1")
self.h2_conv = ConvBNLayer(
in_channels=self.out_channels+self.in_channels[2],
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="unet_h_2")
self.h3_conv = ConvBNLayer(
in_channels=self.out_channels+self.in_channels[3],
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="unet_h_3")
self.g0_deconv = DeConvBNLayer(
in_channels=self.in_channels[0],
out_channels=self.out_channels,
kernel_size=4,
stride=2,
padding=1,
if_act=True,
act='relu',
name="unet_g_0")
self.g1_deconv = DeConvBNLayer(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=4,
stride=2,
padding=1,
if_act=True,
act='relu',
name="unet_g_1")
self.g2_deconv = DeConvBNLayer(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=4,
stride=2,
padding=1,
if_act=True,
act='relu',
name="unet_g_2")
self.g3_conv = ConvBNLayer(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
if_act=True,
act='relu',
name="unet_g_3")
def forward(self, x):
f = x[::-1]
h = f[0]
g = self.g0_deconv(h)
h = paddle.concat([g, f[1]], axis=1)
h = self.h1_conv(h)
g = self.g1_deconv(h)
h = paddle.concat([g, f[2]], axis=1)
h = self.h2_conv(h)
g = self.g2_deconv(h)
h = paddle.concat([g, f[3]], axis=1)
h = self.h3_conv(h)
g = self.g3_conv(h)
return g

284
ppocr/modeling/necks/sast_fpn.py
View File

File diff suppressed because it is too large Load Diff

4
ppocr/postprocess/__init__.py
Unescape View File

@ -24,11 +24,13 @@ __all__ = ['build_post_process']
def build_post_process(config, global_config=None):
from .db_postprocess import DBPostProcess
from .east_postprocess import EASTPostProcess
from .sast_postprocess import SASTPostProcess
from .rec_postprocess import CTCLabelDecode, AttnLabelDecode
from .cls_postprocess import ClsPostProcess
support_dict = [
'DBPostProcess', 'CTCLabelDecode', 'AttnLabelDecode', 'ClsPostProcess'
'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode', 'AttnLabelDecode', 'ClsPostProcess'
]
config = copy.deepcopy(config)

4
ppocr/postprocess/db_postprocess.py
Unescape View File

@ -138,9 +138,9 @@ class DBPostProcess(object):
boxes_batch = []
for batch_index in range(pred.shape[0]):
height, width = shape_list[batch_index]
src_h, src_w, ratio_h, ratio_w = shape_list[batch_index]
boxes, scores = self.boxes_from_bitmap(
pred[batch_index], segmentation[batch_index], width, height)
pred[batch_index], segmentation[batch_index], src_w, src_h)
boxes_batch.append({'points': boxes})
return boxes_batch

141
ppocr/postprocess/east_postprocess.py
Unescape View File

@ -0,0 +1,141 @@
# Copyright (c) 2020 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 numpy as np
from .locality_aware_nms import nms_locality
import cv2
import os
import sys
# __dir__ = os.path.dirname(os.path.abspath(__file__))
# sys.path.append(__dir__)
# sys.path.append(os.path.abspath(os.path.join(__dir__, '..')))
class EASTPostProcess(object):
"""
The post process for EAST.
"""
def __init__(self,
score_thresh=0.8,
cover_thresh=0.1,
nms_thresh=0.2,
**kwargs):
self.score_thresh = score_thresh
self.cover_thresh = cover_thresh
self.nms_thresh = nms_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 restore_rectangle_quad(self, origin, geometry):
"""
Restore rectangle from quadrangle.
"""
# quad
origin_concat = np.concatenate(
(origin, origin, origin, origin), axis=1) # (n, 8)
pred_quads = origin_concat - geometry
pred_quads = pred_quads.reshape((-1, 4, 2)) # (n, 4, 2)
return pred_quads
def detect(self,
score_map,
geo_map,
score_thresh=0.8,
cover_thresh=0.1,
nms_thresh=0.2):
"""
restore text boxes from score map and geo map
"""
score_map = score_map[0]
geo_map = np.swapaxes(geo_map, 1, 0)
geo_map = np.swapaxes(geo_map, 1, 2)
# filter the score map
xy_text = np.argwhere(score_map > score_thresh)
if len(xy_text) == 0:
return []
# sort the text boxes via the y axis
xy_text = xy_text[np.argsort(xy_text[:, 0])]
#restore quad proposals
text_box_restored = self.restore_rectangle_quad(
xy_text[:, ::-1] * 4, geo_map[xy_text[:, 0], xy_text[:, 1], :])
boxes = np.zeros((text_box_restored.shape[0], 9), dtype=np.float32)
boxes[:, :8] = text_box_restored.reshape((-1, 8))
boxes[:, 8] = score_map[xy_text[:, 0], xy_text[:, 1]]
if self.is_python35:
import lanms
boxes = lanms.merge_quadrangle_n9(boxes, nms_thresh)
else:
boxes = nms_locality(boxes.astype(np.float64), nms_thresh)
if boxes.shape[0] == 0:
return []
# Here we filter some low score boxes by the average score map,
# this is different from the orginal paper.
for i, box in enumerate(boxes):
mask = np.zeros_like(score_map, dtype=np.uint8)
cv2.fillPoly(mask, box[:8].reshape(
(-1, 4, 2)).astype(np.int32) // 4, 1)
boxes[i, 8] = cv2.mean(score_map, mask)[0]
boxes = boxes[boxes[:, 8] > cover_thresh]
return boxes
def sort_poly(self, p):
"""
Sort polygons.
"""
min_axis = np.argmin(np.sum(p, axis=1))
p = p[[min_axis, (min_axis + 1) % 4,\
(min_axis + 2) % 4, (min_axis + 3) % 4]]
if abs(p[0, 0] - p[1, 0]) > abs(p[0, 1] - p[1, 1]):
return p
else:
return p[[0, 3, 2, 1]]
def __call__(self, outs_dict, shape_list):
score_list = outs_dict['f_score']
geo_list = outs_dict['f_geo']
img_num = len(shape_list)
dt_boxes_list = []
for ino in range(img_num):
score = score_list[ino].numpy()
geo = geo_list[ino].numpy()
boxes = self.detect(
score_map=score,
geo_map=geo,
score_thresh=self.score_thresh,
cover_thresh=self.cover_thresh,
nms_thresh=self.nms_thresh)
boxes_norm = []
if len(boxes) > 0:
h, w = score.shape[1:]
src_h, src_w, ratio_h, ratio_w = shape_list[ino]
boxes = boxes[:, :8].reshape((-1, 4, 2))
boxes[:, :, 0] /= ratio_w
boxes[:, :, 1] /= ratio_h
for i_box, box in enumerate(boxes):
box = self.sort_poly(box.astype(np.int32))
if np.linalg.norm(box[0] - box[1]) < 5 \
or np.linalg.norm(box[3] - box[0]) < 5:
continue
boxes_norm.append(box)
dt_boxes_list.append({'points': np.array(boxes_norm)})
return dt_boxes_list

199
ppocr/postprocess/locality_aware_nms.py
Unescape View File

@ -0,0 +1,199 @@
"""
Locality aware nms.
"""
import numpy as np
from shapely.geometry import Polygon
def intersection(g, p):
"""
Intersection.
"""
g = Polygon(g[:8].reshape((4, 2)))
p = Polygon(p[:8].reshape((4, 2)))
g = g.buffer(0)
p = p.buffer(0)
if not g.is_valid or not p.is_valid:
return 0
inter = Polygon(g).intersection(Polygon(p)).area
union = g.area + p.area - inter
if union == 0:
return 0
else:
return inter / union
def intersection_iog(g, p):
"""
Intersection_iog.
"""
g = Polygon(g[:8].reshape((4, 2)))
p = Polygon(p[:8].reshape((4, 2)))
if not g.is_valid or not p.is_valid:
return 0
inter = Polygon(g).intersection(Polygon(p)).area
#union = g.area + p.area - inter
union = p.area
if union == 0:
print("p_area is very small")
return 0
else:
return inter / union
def weighted_merge(g, p):
"""
Weighted merge.
"""
g[:8] = (g[8] * g[:8] + p[8] * p[:8]) / (g[8] + p[8])
g[8] = (g[8] + p[8])
return g
def standard_nms(S, thres):
"""
Standard nms.
"""
order = np.argsort(S[:, 8])[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
ovr = np.array([intersection(S[i], S[t]) for t in order[1:]])
inds = np.where(ovr <= thres)[0]
order = order[inds + 1]
return S[keep]
def standard_nms_inds(S, thres):
"""
Standard nms, retun inds.
"""
order = np.argsort(S[:, 8])[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
ovr = np.array([intersection(S[i], S[t]) for t in order[1:]])
inds = np.where(ovr <= thres)[0]
order = order[inds + 1]
return keep
def nms(S, thres):
"""
nms.
"""
order = np.argsort(S[:, 8])[::-1]
keep = []
while order.size > 0:
i = order[0]
keep.append(i)
ovr = np.array([intersection(S[i], S[t]) for t in order[1:]])
inds = np.where(ovr <= thres)[0]
order = order[inds + 1]
return keep
def soft_nms(boxes_in, Nt_thres=0.3, threshold=0.8, sigma=0.5, method=2):
"""
soft_nms
:para boxes_in, N x 9 (coords + score)
:para threshould, eliminate cases min score(0.001)
:para Nt_thres, iou_threshi
:para sigma, gaussian weght
:method, linear or gaussian
"""
boxes = boxes_in.copy()
N = boxes.shape[0]
if N is None or N < 1:
return np.array([])
pos, maxpos = 0, 0
weight = 0.0
inds = np.arange(N)
tbox, sbox = boxes[0].copy(), boxes[0].copy()
for i in range(N):
maxscore = boxes[i, 8]
maxpos = i
tbox = boxes[i].copy()
ti = inds[i]
pos = i + 1
#get max box
while pos < N:
if maxscore < boxes[pos, 8]:
maxscore = boxes[pos, 8]
maxpos = pos
pos = pos + 1
#add max box as a detection
boxes[i, :] = boxes[maxpos, :]
inds[i] = inds[maxpos]
#swap
boxes[maxpos, :] = tbox
inds[maxpos] = ti
tbox = boxes[i].copy()
pos = i + 1
#NMS iteration
while pos < N:
sbox = boxes[pos].copy()
ts_iou_val = intersection(tbox, sbox)
if ts_iou_val > 0:
if method == 1:
if ts_iou_val > Nt_thres:
weight = 1 - ts_iou_val
else:
weight = 1
elif method == 2:
weight = np.exp(-1.0 * ts_iou_val**2 / sigma)
else:
if ts_iou_val > Nt_thres:
weight = 0
else:
weight = 1
boxes[pos, 8] = weight * boxes[pos, 8]
#if box score falls below thresold, discard the box by
#swaping last box update N
if boxes[pos, 8] < threshold:
boxes[pos, :] = boxes[N - 1, :]
inds[pos] = inds[N - 1]
N = N - 1
pos = pos - 1
pos = pos + 1
return boxes[:N]
def nms_locality(polys, thres=0.3):
"""
locality aware nms of EAST
:param polys: a N*9 numpy array. first 8 coordinates, then prob
:return: boxes after nms
"""
S = []
p = None
for g in polys:
if p is not None and intersection(g, p) > thres:
p = weighted_merge(g, p)
else:
if p is not None:
S.append(p)
p = g
if p is not None:
S.append(p)
if len(S) == 0:
return np.array([])
return standard_nms(np.array(S), thres)
if __name__ == '__main__':
# 343,350,448,135,474,143,369,359
print(
Polygon(np.array([[343, 350], [448, 135], [474, 143], [369, 359]]))
.area)

2
ppocr/postprocess/rec_postprocess.py
Unescape View File

@ -27,7 +27,7 @@ class BaseRecLabelDecode(object):
'ch', 'en', 'en_sensitive', 'french', 'german', 'japan', 'korean'
]
assert character_type in support_character_type, "Only {} are supported now but get {}".format(
support_character_type, self.character_str)
support_character_type, character_type)
if character_type == "en":
self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz"

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save