最近在研究文本檢測(cè)網(wǎng)絡(luò)EAST浪感，目前EAST已經(jīng)有最新的AdvancedEAST,先看一下網(wǎng)絡(luò)架構(gòu)。
AdvancedEAST.png
其實(shí)網(wǎng)絡(luò)架構(gòu)挺簡(jiǎn)單的，類(lèi)似于U-net差导，就是在以前EAST網(wǎng)絡(luò)的基礎(chǔ)上被啼，將主網(wǎng)絡(luò)替換為VGG或者ResNet帜消。
發(fā)現(xiàn)網(wǎng)上和Github上大部分實(shí)現(xiàn)是基于Keras或者Tensorflow的，很少有pytorch實(shí)現(xiàn)趟据，于是就自己動(dòng)手用pytorch實(shí)現(xiàn)了一遍券犁。Github有一個(gè)基于keras實(shí)現(xiàn)是用遞歸方法，這樣看起來(lái)代碼很優(yōu)雅汹碱，但是不好理解粘衬。
可參考https://github.com/huoyijie/AdvancedEAST
以下是用pytorch基于VGG16的AdvancedEAST
import torch
import torch.nn as nn
from torchvision.models import VGG
def make_layers(cfg, batch_norm=False):
layers = []
in_channels = 3
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return nn.Sequential(*layers)
# "A": vgg_11, "B": vgg_13, "D": vgg_16, "E": vgg_19
cfgs = {
'A': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'B': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'D': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'],
'E': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'],
}
# 提取VGG模型訓(xùn)練參數(shù)
class Extractor(nn.Module):
def __init__(self, pretrained=False):
super(Extractor, self).__init__()
vgg16_bn = VGG(make_layers(cfgs["D"], batch_norm=True))
if pretrained:
vgg16_bn.load_state_dict(torch.load('./pths/vgg16_bn-6c64b313.pth'))
self.features = vgg16_bn.features
def forward(self, x):
out = []
for m in self.features:
x = m(x)
# 提取maxpool層為后續(xù)合并
if isinstance(m, nn.MaxPool2d):
out.append(x)
return out[1:]
class UpSampling(nn.Module):
def __init__(self, in_channels, out_channels):
super(UpSampling, self).__init__()
self.layer = nn.Sequential(
nn.ConvTranspose2d(in_channels, out_channels, 3, 2, 1, 1)
)
def forward(self, x):
return self.layer(x)
class Conv_BN_ReLU(nn.Module):
def __init__(self, in_channels, out_channels):
super(Conv_BN_ReLU, self).__init__()
self.con_bn_relu = nn.Sequential(
nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
)
def forward(self, x):
return self.con_bn_relu(x)
class FPN_16(nn.Module):
def __init__(self):
super(FPN_16, self).__init__()
self.extractor = Extractor()
self.bn = nn.BatchNorm2d(512)
self.upsampling1 = UpSampling(512, 256)
self.conv1 = Conv_BN_ReLU(768, 128)
self.upsampling2 = UpSampling(128, 64)
self.conv2 = Conv_BN_ReLU(320, 64)
self.upsampling3 = UpSampling(64, 32)
self.conv3 = Conv_BN_ReLU(160, 32)
self.conv4 = nn.Conv2d(32, 32, 3, 1, 1)
def forward(self, inputs):
f4, f3, f2, f1 = self.extractor(inputs)
h1 = self.bn(f1)
h2 = self.conv1(torch.cat((self.upsampling1(h1), f2), 1))
h3 = self.conv2(torch.cat((self.upsampling2(h2), f3), 1))
h4 = self.conv3(torch.cat((self.upsampling3(h3), f4), 1))
out = self.conv4(h4)
return out
class EAST(nn.Module):
def __init__(self):
super(EAST, self).__init__()
self.fpn = FPN_16()
self.inside_score = nn.Conv2d(32, 1, 1)
self.side_vertex_code = nn.Conv2d(32, 2, 1)
self.side_vertex_geo = nn.Conv2d(32, 4, 1)
def forward(self, input):
fpn = self.fpn(input)
ins = self.inside_score(fpn)
svc = self.side_vertex_code(fpn)
svg = self.side_vertex_geo(fpn)
out = torch.cat((ins, svc, svg), 1)
return out
if __name__ == "__main__":
model = EAST().cuda()
x = torch.randn(2, 3, 224, 224).cuda()
out = model(x)
print(out.shape)
