DeeplabV3+解码器复现（二）

1.全局卷积网络模块

全局卷积网络模块(Global Convolutional Network, GCN)是由清华大学旷视科技提出应用于语义分割任务改善模型性能的模块，其核心是利用非对称卷积获得大感受野来解决语义分割任务中像素的分类和定位问题，以达到更好的分割效果。相较于普通卷积，非对称卷积能够在少参数量的情况下达到大卷积核的特征映射效果。以一个3×3的普通卷积为例，其卷积的特征映射等于3×1和1×3两个非对称卷积的卷积结果。GCN模块由1×k+k×1和k×1+1×k两组非对称卷积构成，并通过稠密连接特征映射出大小为k×k的特征图，最后相加融合输出预测图像，其中k为卷积核大小，n为预测类别数(包含背景)。

    self.branch1_0 = nn.Sequential(nn.Conv2d(24, 4, (15, 1), 1, (7, 0)), nn.Conv2d(4, 4, (1, 15), 1, (0, 7))) self.branch1_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 15), 1, (0, 7)), nn.Conv2d(4, 4, (15, 1), 1, (7, 0)))
def forward(self, x):branch1_0 = self.branch1_0(x)branch1_1 = self.branch1_1(x)branch1 = branch1_0 + branch1_1

2.边界细化模块

边界细化模块(Boundary Refinement，BR)是在全局卷积网络中提出的边界优化模块，其结构设计类似于残差结构，通过短接对预测结果进行进一步修正，从而提高目标对象的定位，达到优化边界的作用，BR模块由两个3×3卷积通过映射短接构成，将残差结构结果与原特征进行相加融合后，再输出边界增强后的预测结果。

self.br = nn.Sequential(nn.Conv2d(4, 4, 3, 1, 1),nn.ReLU(),nn.Conv2d(4, 4, 3, 1, 1))

3.解码器设计one

    def forward(self, x):H, W = x.size(2), x.size(3)the_two_features, low_level_features, the_three_features, the_four_features, x = self.backbone(x)x = self.aspp(x) #32*32*256x0 = F.interpolate(x, size=(low_level_features.size(2), low_level_features.size(3)),mode='bilinear', align_corners=True)  #32*32*256-128*128*256x1 = torch.cat((x0, low_level_features), dim=1)  #128*128*280x2 = self.conv3(x1) #128*128*4# -----------------------------------------## 输入128*128*24，输出128*128*4branch1_0 = self.branch1_0(low_level_features)branch1_1 = self.branch1_1(low_level_features)branch1 = branch1_0 + branch1_1branch = self.br(branch1)result1 = branch + branch1 #128*128*4# -----------------------------------------#x3 = x2 + result1 #128*128*4x4 = self.br(x3) #128*128*4x5 = F.interpolate(x4, size=(the_two_features.size(2), the_two_features.size(3)),mode='bilinear', align_corners=True)  #128*128*4-256*256*4# -----------------------------------------## 输入256*256*16，输出256*256*4branch2_0 = self.branch2_0(the_two_features)branch2_1 = self.branch2_1(the_two_features)branch2 = branch2_0 + branch2_1branchh = self.br(branch2)result2 = branchh + branch2  #256*256*4# -----------------------------------------#x6 = x5 + result2  #256*256*4x7 = self.br(x6)  #256*256*4x8 = F.interpolate(x7, size=(H, W), mode='bilinear', align_corners=True)return x8

3.解码器设计two

self.branch1_0 = nn.Sequential(nn.Conv2d(24, 4, (3, 1), 1, (1, 0)), nn.Conv2d(4, 4, (1, 3), 1, (0, 1))) self.branch1_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 3), 1, (0, 1)), nn.Conv2d(4, 4, (3, 1), 1, (1, 0))) self.branch2_0 = nn.Sequential(nn.Conv2d(24, 4, (7, 1), 1, (3, 0)),  nn.Conv2d(4, 4, (1, 7), 1, (0, 3))) self.branch2_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 7), 1, (0, 3)),nn.Conv2d(4, 4, (7, 1), 1, (3, 0)))  self.branch3_0 = nn.Sequential(nn.Conv2d(24, 4, (11, 1), 1, (5, 0)),nn.Conv2d(4, 4, (1, 11), 1, (0, 5))) self.branch3_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 11), 1, (0, 5)), nn.Conv2d(4, 4, (11, 1), 1, (5, 0))) self.branch4_0 = nn.Sequential(nn.Conv2d(24, 4, (15, 1), 1, (7, 0)), nn.Conv2d(4, 4, (1, 15), 1, (0, 7))) self.branch4_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 15), 1, (0, 7)), nn.Conv2d(4, 4, (15, 1), 1, (7, 0))) self.br = nn.Sequential(nn.Conv2d(4, 4, 3, 1, 1),nn.ReLU(),nn.Conv2d(4, 4, 3, 1, 1))

        self.cat_conv = nn.Sequential(nn.Conv2d(296, 128, 3, stride=1, padding=1),nn.BatchNorm2d(128),nn.ReLU(inplace=True),nn.Dropout(0.5),nn.Conv2d(128, 4, 3, stride=1, padding=1),nn.BatchNorm2d(4),nn.ReLU(inplace=True),nn.Dropout(0.1),)self.branch1_0 = nn.Sequential(nn.Conv2d(24, 4, (3, 1), 1, (1, 0)), nn.Conv2d(4, 4, (1, 3), 1, (0, 1))) self.branch1_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 3), 1, (0, 1)), nn.Conv2d(4, 4, (3, 1), 1, (1, 0))) self.branch2_0 = nn.Sequential(nn.Conv2d(24, 4, (7, 1), 1, (3, 0)),  nn.Conv2d(4, 4, (1, 7), 1, (0, 3)))  self.branch2_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 7), 1, (0, 3)),  nn.Conv2d(4, 4, (7, 1), 1, (3, 0)))  self.branch3_0 = nn.Sequential(nn.Conv2d(24, 4, (11, 1), 1, (5, 0)),  nn.Conv2d(4, 4, (1, 11), 1, (0, 5)))  self.branch3_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 11), 1, (0, 5)),  nn.Conv2d(4, 4, (11, 1), 1, (5, 0)))  self.branch4_0 = nn.Sequential(nn.Conv2d(24, 4, (15, 1), 1, (7, 0)),  nn.Conv2d(4, 4, (1, 15), 1, (0, 7)))  self.branch4_1 = nn.Sequential(nn.Conv2d(24, 4, (1, 15), 1, (0, 7)), nn.Conv2d(4, 4, (15, 1), 1, (7, 0)))  self.br = nn.Sequential(nn.Conv2d(4, 4, 3, 1, 1),nn.ReLU(),nn.Conv2d(4, 4, 3, 1, 1))def forward(self, x):H, W = x.size(2), x.size(3)# -----------------------------------------##   获得两个特征层#   low_level_features: 浅层特征-进行卷积处理#   x : 主干部分-利用ASPP结构进行加强特征提取# -----------------------------------------#low_level_features, x = self.backbone(x)x = self.aspp(x) #32*32*256x0 = F.interpolate(x, size=(low_level_features.size(2), low_level_features.size(3)),mode='bilinear', align_corners=True)  #32*32*256-128*128*256# -----------------------------------------## 输入128*128*24，输出128*128*4branch1_0 = self.branch1_0(low_level_features)branch1_1 = self.branch1_1(low_level_features)branch1 = branch1_0 + branch1_1br1 = self.br(branch1)result1 = br1 + branch1  # 128*128*4# -----------------------------------------## 输入128*128*24，输出128*128*4branch2_0 = self.branch2_0(low_level_features)branch2_1 = self.branch2_1(low_level_features)branch2 = branch2_0 + branch2_1br2 = self.br(branch2)result2 = br2 + branch2  # 128*128*4# -----------------------------------------## 输入128*128*24，输出128*128*4branch3_0 = self.branch3_0(low_level_features)branch3_1 = self.branch3_1(low_level_features)branch3 = branch3_0 + branch3_1br3= self.br(branch3)result3 = br3 + branch3  # 128*128*4# -----------------------------------------## 输入128*128*24，输出128*128*4branch4_0 = self.branch4_0(low_level_features)branch4_1 = self.branch4_1(low_level_features)branch4 = branch4_0 + branch4_1br4 = self.br(branch4)result4 = br4 + branch4  # 128*128*4# -----------------------------------------#x1 = torch.cat((result1, result2, result3, result4), dim=1)  # 128*128*16x2 = torch.cat((x1, low_level_features, x0), dim=1)  # 128*128*296x3 = self.cat_conv(x2) # 128*128*296-128*128*4x4 = F.interpolate(x3, size=(H, W), mode='bilinear', align_corners=True) # 128*128*4-512*512*4return x4

DeeplabV3+解码器复现（二）相关推荐

DeeplabV3+解码器复现
1.全局通道注意力全局通道注意力结构可以使神经网络在训练过程中建模特征图各个通道的重要性,自动判断哪些通道上的信息为有用信息,哪些通道上的信息为噪声信息.全局通道注意力结构以SENet的通道注意力结 ...
DIY个人智能家庭网关——硬件篇之433M无线接收解码器（二）
为了降低DIY难度,我提供了方案二:采用现成的模块来搭建解码器,需要如下模块 1.STM8S103F单片机模块和编程器 2.433M无线接收模块 3.PL2303 USB转串口模块或带线的另外还需 ...
FFplay文档解读-5-编解码器选项二
设置sub pel me比较功能. 可能的取值如下: 值解释 sad 绝对差值之和,fast(默认) sse 平方误差之和 satd 绝对Hadamard的总和改变了差异 dct 绝对DCT转换差异 ...
语义分割之Deeplabv3源码解读
写在前面,本文的代码是基于Pytorch框架resnet101基础网络的DeepLabv3网络复现先贴一张代码效果图: 预备知识本文提到的output_stride来源于deeplabv3论文,指 ...
java struts2 漏洞复现合集
目录一.S2-001复现二.S2-005复现三.S2-007复现四.S2-008复现五.S2-009复现六.S2-012复现七.S2-013复现八.S2-015复现九.S2-016复 ...
【转载】裸眼识别二维码
一个开关的开合状态可以对应"0.1"两个信号,莱布尼茨三百多年前留下的这个制式影响深远.日常语言依赖于思维逻辑,与之类似,数字语言的流通则依赖于数学逻辑,但本质上仍然只是无数个与非 ...
帝国cms7.5漏洞复现
EmpireCMS v7.5后台任意代码执行分析目录一.漏洞简介二.漏洞复现三.原理分析一.漏洞简介 EmpireCMS 7.5版本及之前版本在后台备份数据库时,未对数据库表名做验证,通过修 ...
编码器-解码器入门级理解
编码器-解码器编码器-解码器使用场景一.PCA 二.自编码器(AE) 1.欠完备自编码器 2.过完备自编码器 3.正则化自编码器 ①.去噪自编码器 ②.稀疏自编码器三.变分自编码器(VAE) 编 ...
【Pytorch项目实战】之生成式网络：编码器-解码器、自编码器AE、变分自编码器VAE、生成式对抗网络GAN
文章目录博主精品专栏导航生成式网络 - 生成合成图像算法一:编码器-解码器算法二:自编码器(Auto-Encoder,AE) 算法三:变分自编码器(Variational Auto Encod ...