mirrors/PaddleOCR
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/PaddleOCR.git synced 2025-06-03 21:53:39 +08:00
Code Issues Packages Projects Releases Wiki Activity

add db++

Browse Source
This commit is contained in:
wangjingyeye 2022-06-27 09:43:14 +00:00
parent 26a89db76a
commit 78064ad956
3 changed files with 33 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ppocr/data/imaug/operators.py
View File

@ -242,8 +242,8 @@ class DetResizeForTest(object):
if 'image_shape' in kwargs: if 'image_shape' in kwargs:
self.image_shape = kwargs['image_shape'] self.image_shape = kwargs['image_shape']
self.resize_type = 1 self.resize_type = 1
if 'keep_ratio' in kwargs: ###### if 'keep_ratio' in kwargs:
self.keep_ratio = kwargs['keep_ratio'] ####### self.keep_ratio = kwargs['keep_ratio']
elif 'limit_side_len' in kwargs: elif 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len'] self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min') self.limit_type = kwargs.get('limit_type', 'min')
@ -273,7 +273,7 @@ class DetResizeForTest(object):
def resize_image_type1(self, img): def resize_image_type1(self, img):
resize_h, resize_w = self.image_shape resize_h, resize_w = self.image_shape
ori_h, ori_w = img.shape[:2] # (h, w, c) ori_h, ori_w = img.shape[:2] # (h, w, c)
if self.keep_ratio: ######## if self.keep_ratio:
resize_w = ori_w * resize_h / ori_h resize_w = ori_w * resize_h / ori_h
N = math.ceil(resize_w / 32) N = math.ceil(resize_w / 32)
resize_w = N * 32 resize_w = N * 32

2
ppocr/modeling/backbones/det_resnet.py
View File

@ -1,4 +1,4 @@
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. # copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
# #
# Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.

47
ppocr/modeling/necks/db_fpn.py
View File

@ -105,7 +105,7 @@ class DSConv(nn.Layer):
class DBFPN(nn.Layer): class DBFPN(nn.Layer):
def __init__(self, in_channels, out_channels, use_asf=None, **kwargs): def __init__(self, in_channels, out_channels, use_asf=False, **kwargs):
super(DBFPN, self).__init__() super(DBFPN, self).__init__()
self.out_channels = out_channels self.out_channels = out_channels
self.use_asf = use_asf self.use_asf = use_asf
@ -164,7 +164,7 @@ class DBFPN(nn.Layer):
weight_attr=ParamAttr(initializer=weight_attr), weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False) bias_attr=False)
if self.use_asf: if self.use_asf is True:
self.asf = ASFBlock(self.out_channels, self.out_channels // 4) self.asf = ASFBlock(self.out_channels, self.out_channels // 4)
def forward(self, x): def forward(self, x):
@ -192,7 +192,7 @@ class DBFPN(nn.Layer):
fuse = paddle.concat([p5, p4, p3, p2], axis=1) fuse = paddle.concat([p5, p4, p3, p2], axis=1)
if self.use_asf: if self.use_asf is True:
fuse = self.asf(fuse, [p5, p4, p3, p2]) fuse = self.asf(fuse, [p5, p4, p3, p2])
return fuse return fuse
@ -367,7 +367,19 @@ class LKPAN(nn.Layer):
class ASFBlock(nn.Layer): class ASFBlock(nn.Layer):
"""
This code is refered from:
https://github.com/MhLiao/DB/blob/master/decoders/feature_attention.py
"""
def __init__(self, in_channels, inter_channels, out_features_num=4): def __init__(self, in_channels, inter_channels, out_features_num=4):
"""
Adaptive Scale Fusion (ASF) block of DBNet++
Args:
in_channels: the number of channels in the input data
inter_channels: the number of middle channels
out_features_num: the number of fused stages
"""
super(ASFBlock, self).__init__() super(ASFBlock, self).__init__()
weight_attr = paddle.nn.initializer.KaimingUniform() weight_attr = paddle.nn.initializer.KaimingUniform()
self.in_channels = in_channels self.in_channels = in_channels
@ -375,39 +387,38 @@ class ASFBlock(nn.Layer):
self.out_features_num = out_features_num self.out_features_num = out_features_num
self.conv = nn.Conv2D(in_channels, inter_channels, 3, padding=1) self.conv = nn.Conv2D(in_channels, inter_channels, 3, padding=1)
self.attention_block_1 = nn.Sequential( self.spatial_scale = nn.Sequential(
#Nx1xHxW #Nx1xHxW
nn.Conv2D( nn.Conv2D(
1, in_channels=1,
1, out_channels=1,
3, kernel_size=3,
bias_attr=False, bias_attr=False,
padding=1, padding=1,
weight_attr=ParamAttr(initializer=weight_attr)), weight_attr=ParamAttr(initializer=weight_attr)),
nn.ReLU(), nn.ReLU(),
nn.Conv2D( nn.Conv2D(
1, in_channels=1,
1, out_channels=1,
1, kernel_size=1,
bias_attr=False, bias_attr=False,
weight_attr=ParamAttr(initializer=weight_attr)), weight_attr=ParamAttr(initializer=weight_attr)),
nn.Sigmoid()) nn.Sigmoid())
self.attention_block_2 = nn.Sequential( self.channel_scale = nn.Sequential(
nn.Conv2D( nn.Conv2D(
inter_channels, in_channels=inter_channels,
out_features_num, out_channels=out_features_num,
1, kernel_size=1,
bias_attr=False, bias_attr=False,
weight_attr=ParamAttr(initializer=weight_attr)), weight_attr=ParamAttr(initializer=weight_attr)),
nn.Sigmoid()) nn.Sigmoid())
def forward(self, fuse_features, features_list): def forward(self, fuse_features, features_list):
fuse_features = self.conv(fuse_features) fuse_features = self.conv(fuse_features)
attention_scores = self.attention_block_1( spatial_x = paddle.mean(fuse_features, axis=1, keepdim=True)
paddle.mean( attention_scores = self.spatial_scale(spatial_x) + fuse_features
fuse_features, axis=1, keepdim=True)) + fuse_features attention_scores = self.channel_scale(attention_scores)
attention_scores = self.attention_block_2(attention_scores)
assert len(features_list) == self.out_features_num assert len(features_list) == self.out_features_num
out_list = [] out_list = []