devbox
繁依Fanyi0
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

基于MobileNetV2主干的DeeplabV3+语义分割实现_mobilenet+deeplabv3+

作者：繁依Fanyi0 | 2024-06-05 21:33:05

赞

踩

mobilenet+deeplabv3+

目录

一. 语义分割的含义

二. DeepLabV3+ 模型

三. 模型整体框架

四. 模型检测效果

五. 代码实现

 六. 源码地址

一. 语义分割的含义

        语义分割是计算机视觉中的基本任务，在语义分割中我们需要将视觉输入分为不同的语义可解释类别，「语义的可解释性」即分类类别在真实世界中是有意义的。例如，我们可能需要区分图像中属于汽车的所有像素，并把这些像素涂成蓝色。

我们将   图像分类，目标检测  和   语义分割 进行对比 可以让我们更好的理解语义分割。
 图像分类：  通过  提取特征，输出 待测图片趋向于某个种类
 目标检测：   通过  提取特征，输出 待测图片中不同物体的位置与种类 
 语义分割：  通过 提取特征， 输出 待测图片的每个像素点的种类

二. DeepLabV3+ 模型

 如上图， Encoder中DCNN部分代表语义分割中的主干网络， 在本文中为轻量网络MobileNetV2
 特征提取分为高层语义提取和低层的语义提取两个部分。
        首先 1 x 1 对通道上关联，起了一个全连接的作用，接下来是 3 个空洞卷积，有关空洞卷积参见。pooling ，然后经过 concate 将这些特征图进行组合，随后经过 1x1 卷积来改变通道大小。接下里对于底层特征图首先进行 1x1 卷积进行通道变换，这样可以拿到一些低层特征，在将上面组合变换通道数的特征图进行一次 4 倍上采样得到和低层特征图大小相同特征图后，进行组合后再进行一次 4 倍上采样。

三. 模型整体框架

        自2017年mobile net问世之后，研究人员就不断在追求更小，更快，更准的网络模型。在这个过程中，也发现了mobile net存在的问题：1.1.结构简单，mobile net使用类似VGG的结构，这种结构已经被证明不如resnet bottle neck结构；2.depthwise convolution的输出，在relu的作用下，很容易废掉。即输出为0，且无法恢复。针对上述问题，谷歌做了改进，也即是本文的主角，mobile net v2。mobile net v2的主要改进为引入了Inverted residual block和利用线性变换替换relu。

四. 模型检测效果

五. 代码实现

MobileV2 网络代码搭建

  1. import math
  2. import os
  3.  
  4. import torch
  5. import torch.nn as nn
  6. import torch.utils.model_zoo as model_zoo
  7.  
  8. BatchNorm2d = nn.BatchNorm2d
  9.  
  10. # PW、DW -> https://blog.csdn.net/qq_41895003/article/details/107408390
  11. # MobileNet V1、V2、V3 -> https://www.icode9.com/content-4-891085.html
  12.  
  13.  
  14. def conv_bn(inp, oup, stride):
  15.     return nn.Sequential(
  16.         nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
  17.         BatchNorm2d(oup),
  18.         nn.ReLU6(inplace=True)
  19.     )
  20.  
  21. def conv_1x1_bn(inp, oup):
  22.     return nn.Sequential(
  23.         nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
  24.         BatchNorm2d(oup),
  25.         nn.ReLU6(inplace=True)
  26.     )
  27.  
  28. # 深度可分离卷积(Depthwise Separable Convolution)
  29. # 一层深度卷积(Depthwise Convolution，DW）与一层逐点卷积（Pointwise Convolution，PW)组合
  30.  
  31. # 倒残差结构Block   PW升维 -> DW -> PW降维
  32. # 在 深度可分离卷积(DW + PW降维) 前加一层 PW
  33. # rate为卷积膨胀系数 若rate>1 则为膨胀卷积(空洞卷积)
  34. # nn.Conv2d(in_channels, out_channels, kernel_size, stride=1,padding=0, dilation=1, groups=1,bias=True):
  35. class InvertedResidual(nn.Module):
  36.     def __init__(self, inp, oup, stride, expand_ratio):
  37.         super(InvertedResidual, self).__init__()
  38.         self.stride = stride
  39.         assert stride in [1, 2]    # assert in 断言, 若stride不在[1, 2]中则报错
  40.  
  41.         hidden_dim = round(inp * expand_ratio)
  42.         self.use_res_connect = self.stride == 1 and inp == oup
  43.  
  44.         # --------------------------------------------#
  45.         #  深度可分离卷积
  46.         #  第一部分：DW, groups = 输出通道数 = 输入通道数, 当group = 1 时 即为普通卷积
  47.         #  第二部分：PW, 利用1×1的卷积更改输出通道数
  48.         # --------------------------------------------#
  49.         if expand_ratio == 1:
  50.             self.conv = nn.Sequential(
  51.                 #--------------------------------------------#
  52.                 #   进行3x3的逐层卷积，进行跨特征点的特征提取
  53.                 #--------------------------------------------#
  54.                 nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
  55.                 BatchNorm2d(hidden_dim),
  56.                 nn.ReLU6(inplace=True),
  57.                 #-----------------------------------#
  58.                 #   利用1x1卷积进行通道数的调整
  59.                 #-----------------------------------#
  60.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
  61.                 BatchNorm2d(oup),
  62.             )
  63.         else:
  64.             self.conv = nn.Sequential(
  65.                 #-----------------------------------#
  66.                 #   利用1x1卷积进行通道数的上升
  67.                 #-----------------------------------#
  68.                 nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
  69.                 BatchNorm2d(hidden_dim),
  70.                 nn.ReLU6(inplace=True),
  71.                 #--------------------------------------------#
  72.                 #   进行3x3的逐层卷积，进行跨特征点的特征提取
  73.                 #--------------------------------------------#
  74.                 nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
  75.                 BatchNorm2d(hidden_dim),
  76.                 nn.ReLU6(inplace=True),
  77.                 #-----------------------------------#
  78.                 #   利用1x1卷积进行通道数的下降
  79.                 #-----------------------------------#
  80.                 nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
  81.                 BatchNorm2d(oup),
  82.             )
  83.  
  84.     def forward(self, x):
  85.         if self.use_res_connect:
  86.             return x + self.conv(x)
  87.         else:
  88.             return self.conv(x)
  89.  
  90. class MobileNetV2(nn.Module):
  91.     def __init__(self, n_class=1000, input_size=224, width_mult=1.):
  92.         super(MobileNetV2, self).__init__()
  93.         block = InvertedResidual
  94.         input_channel = 32
  95.         last_channel = 1280
  96.         interverted_residual_setting = [
  97.             # t, c, n, s
  98.             [1, 16, 1, 1], # 256, 256, 32 -> 256, 256, 16
  99.             [6, 24, 2, 2], # 256, 256, 16 -> 128, 128, 24   2
  100.             [6, 32, 3, 2], # 128, 128, 24 -> 64, 64, 32     4
  101.             [6, 64, 4, 2], # 64, 64, 32 -> 32, 32, 64       7
  102.             [6, 96, 3, 1], # 32, 32, 64 -> 32, 32, 96
  103.             [6, 160, 3, 2], # 32, 32, 96 -> 16, 16, 160     14
  104.             [6, 320, 1, 1], # 16, 16, 160 -> 16, 16, 320
  105.         ]
  106.  
  107.         assert input_size % 32 == 0
  108.         input_channel = int(input_channel * width_mult)
  109.         self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
  110.  
  111.         # 512, 512, 3 -> 256, 256, 32
  112.         # 对应 nets/nets.jpg中的MobilenetV2表中的第一个Conv2d
  113.         self.features = [conv_bn(3, input_channel, 2)]
  114.  
  115.         for t, c, n, s in interverted_residual_setting:
  116.             output_channel = int(c * width_mult)
  117.             # 每一个blocks中包括 n个残差block, 第一个block的步长为s, 剩下的为1
  118.             for i in range(n):
  119.                 if i == 0:
  120.                     self.features.append(block(input_channel, output_channel, s, expand_ratio=t))
  121.                 else:
  122.                     self.features.append(block(input_channel, output_channel, 1, expand_ratio=t))
  123.                 input_channel = output_channel
  124.  
  125.         self.features.append(conv_1x1_bn(input_channel, self.last_channel))
  126.         self.features = nn.Sequential(*self.features)
  127.  
  128.         self.classifier = nn.Sequential(
  129.             nn.Dropout(0.2),
  130.             nn.Linear(self.last_channel, n_class),
  131.         )
  132.  
  133.         self._initialize_weights()
  134.  
  135.     def forward(self, x):
  136.         x = self.features(x)
  137.         x = x.mean(3).mean(2)
  138.         x = self.classifier(x)
  139.         return x
  140.  
  141.     # isinstance(x, y)判断x , y是否时相同类型 ，返回bool类型
  142.     # 例如：设置一个条件,如果m为Conv2d层就为该m添加相应的参数
  143.     def _initialize_weights(self):
  144.         for m in self.modules():
  145.             if isinstance(m, nn.Conv2d):
  146.                 n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
  147.                 m.weight.data.normal_(0, math.sqrt(2. / n))
  148.                 if m.bias is not None:
  149.                     m.bias.data.zero_()
  150.             elif isinstance(m, BatchNorm2d):
  151.                 m.weight.data.fill_(1)
  152.                 m.bias.data.zero_()
  153.             elif isinstance(m, nn.Linear):
  154.                 n = m.weight.size(1)
  155.                 m.weight.data.normal_(0, 0.01)
  156.                 m.bias.data.zero_()
  157.  
  158.  
  159. def load_url(url, model_dir='./model_data', map_location=None):
  160.     if not os.path.exists(model_dir):
  161.         os.makedirs(model_dir)
  162.     filename = url.split('/')[-1]
  163.     cached_file = os.path.join(model_dir, filename)
  164.     if os.path.exists(cached_file):
  165.         return torch.load(cached_file, map_location=map_location)
  166.     else:
  167.         return model_zoo.load_url(url,model_dir=model_dir)
  168.  
  169. def mobilenetv2(pretrained=False, **kwargs):
  170.     model = MobileNetV2(n_class=1000, **kwargs)
  171.     if pretrained:
  172.         model.load_state_dict(load_url('https://github.com/bubbliiiing/deeplabv3-plus-pytorch/releases/download/v1.0/mobilenet_v2.pth.tar'), strict=False)
  173.     return model
  174.  
  175. if __name__ == "__main__":
  176.     model = mobilenetv2()
  177.     for i, layer in enumerate(model.features):
  178.         print(i, layer)

DeepLabV3 + 网络代码搭建

  1. import torch
  2. import torch.nn as nn
  3. import torch.nn.functional as F
  4. from nets.xception import xception
  5. from nets.mobilenetV2 import mobilenetv2
  6.  
  7.  
  8. class MobileNetV2(nn.Module):
  9.     def __init__(self, downsample_factor=8, pretrained=True):
  10.         super(MobileNetV2, self).__init__()
  11.         from functools import partial
  12.  
  13.         model = mobilenetv2(pretrained)
  14.         # res = [0, 1, 2, 3, 4]
  15.         # print(res[:-1])
  16.         # out:[0, 1, 2, 3]
  17.         self.features = model.features[:-1]
  18.  
  19.         # [2, 4, 7, 14]  代表的是  self.features 中层的位置
  20.         self.total_idx = len(self.features)
  21.         self.down_idx = [2, 4, 7, 14]
  22.  
  23.         if downsample_factor == 8:
  24.             for i in range(self.down_idx[-2], self.down_idx[-1]):
  25.                 self.features[i].apply(
  26.                     partial(self._nostride_dilate, dilate=2)
  27.                 )
  28.             for i in range(self.down_idx[-1], self.total_idx):
  29.                 self.features[i].apply(
  30.                     partial(self._nostride_dilate, dilate=4)
  31.                 )
  32.         elif downsample_factor == 16:
  33.             for i in range(self.down_idx[-1], self.total_idx):
  34.                 self.features[i].apply(
  35.                     partial(self._nostride_dilate, dilate=2)
  36.                 )
  37.  
  38.     #  dilate 膨胀系数
  39.     def _nostride_dilate(self, m, dilate):
  40.         classname = m.__class__.__name__
  41.         if classname.find('Conv') != -1:
  42.             if m.stride == (2, 2):
  43.                 m.stride = (1, 1)
  44.                 if m.kernel_size == (3, 3):
  45.                     m.dilation = (dilate // 2, dilate // 2)
  46.                     m.padding = (dilate // 2, dilate // 2)
  47.             else:
  48.                 if m.kernel_size == (3, 3):
  49.                     m.dilation = (dilate, dilate)
  50.                     m.padding = (dilate, dilate)
  51.  
  52.     def forward(self, x):
  53.         low_level_features = self.features[:4](x)
  54.         x = self.features[4:](low_level_features)
  55.         return low_level_features, x
  56.  
  57.     # -----------------------------------------#
  58.  
  59.  
  60. #   ASPP特征提取模块
  61. #   利用不同膨胀率的膨胀卷积进行特征提取
  62. # -----------------------------------------#
  63. class ASPP(nn.Module):
  64.     def __init__(self, dim_in, dim_out, rate=1, bn_mom=0.1):
  65.         super(ASPP, self).__init__()
  66.         self.branch1 = nn.Sequential(
  67.             nn.Conv2d(dim_in, dim_out, 1, 1, padding=0, dilation=rate, bias=True),
  68.             nn.BatchNorm2d(dim_out, momentum=bn_mom),
  69.             nn.ReLU(inplace=True),
  70.         )
  71.         self.branch2 = nn.Sequential(
  72.             nn.Conv2d(dim_in, dim_out, 3, 1, padding=6 * rate, dilation=6 * rate, bias=True),
  73.             nn.BatchNorm2d(dim_out, momentum=bn_mom),
  74.             nn.ReLU(inplace=True),
  75.         )
  76.         self.branch3 = nn.Sequential(
  77.             nn.Conv2d(dim_in, dim_out, 3, 1, padding=12 * rate, dilation=12 * rate, bias=True),
  78.             nn.BatchNorm2d(dim_out, momentum=bn_mom),
  79.             nn.ReLU(inplace=True),
  80.         )
  81.         self.branch4 = nn.Sequential(
  82.             nn.Conv2d(dim_in, dim_out, 3, 1, padding=18 * rate, dilation=18 * rate, bias=True),
  83.             nn.BatchNorm2d(dim_out, momentum=bn_mom),
  84.             nn.ReLU(inplace=True),
  85.         )
  86.         self.branch5_conv = nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=True)
  87.         self.branch5_bn = nn.BatchNorm2d(dim_out, momentum=bn_mom)
  88.         self.branch5_relu = nn.ReLU(inplace=True)
  89.  
  90.         self.conv_cat = nn.Sequential(
  91.             nn.Conv2d(dim_out * 5, dim_out, 1, 1, padding=0, bias=True),
  92.             nn.BatchNorm2d(dim_out, momentum=bn_mom),
  93.             nn.ReLU(inplace=True),
  94.         )
  95.  
  96.     def forward(self, x):
  97.         [b, c, row, col] = x.size()
  98.         # -----------------------------------------#
  99.         #   一共五个分支
  100.         # -----------------------------------------#
  101.         conv1x1 = self.branch1(x)
  102.         conv3x3_1 = self.branch2(x)
  103.         conv3x3_2 = self.branch3(x)
  104.         conv3x3_3 = self.branch4(x)
  105.         # -----------------------------------------#
  106.         #   第五个分支，全局平均池化+卷积
  107.         # -----------------------------------------#
  108.         global_feature = torch.mean(x, 2, True)
  109.         global_feature = torch.mean(global_feature, 3, True)
  110.         global_feature = self.branch5_conv(global_feature)
  111.         global_feature = self.branch5_bn(global_feature)
  112.         global_feature = self.branch5_relu(global_feature)
  113.         global_feature = F.interpolate(global_feature, (row, col), None, 'bilinear', True)
  114.  
  115.         # -----------------------------------------#
  116.         #   将五个分支的内容堆叠起来
  117.         #   然后1x1卷积整合特征
  118.         # -----------------------------------------#
  119.         feature_cat = torch.cat([conv1x1, conv3x3_1, conv3x3_2, conv3x3_3, global_feature], dim=1)
  120.  
  121.         # 对应 nets.jpg中 encoder 右侧的 1x1 Covn
  122.         # 利用1x1卷积调整通道数
  123.         # 52, 52, 1280 -> 52,52,256
  124.         result = self.conv_cat(feature_cat)
  125.         return result
  126.  
  127.  
  128. class DeepLab(nn.Module):
  129.     def __init__(self, num_classes, backbone="mobilenet", pretrained=False, downsample_factor=16):
  130.         super(DeepLab, self).__init__()
  131.         if backbone == "xception":
  132.             # ----------------------------------#
  133.             #   获得两个特征层
  134.             #   浅层特征    [128,128,256]
  135.             #   主干部分    [30,30,2048]
  136.             # ----------------------------------#
  137.             self.backbone = xception(downsample_factor=downsample_factor, pretrained=pretrained)
  138.             in_channels = 2048
  139.             low_level_channels = 256
  140.         elif backbone == "mobilenet":
  141.             # ----------------------------------#
  142.             #   获得两个特征层
  143.             #   浅层特征    [128,128,24]
  144.             #   主干部分    [30,30,320]
  145.             # ----------------------------------#
  146.             self.backbone = MobileNetV2(downsample_factor=downsample_factor, pretrained=pretrained)
  147.             in_channels = 320
  148.             low_level_channels = 24
  149.         else:
  150.             raise ValueError('Unsupported backbone - `{}`, Use mobilenet, xception.'.format(backbone))
  151.  
  152.         # -----------------------------------------#
  153.         #   ASPP特征提取模块
  154.         #   利用不同膨胀率的膨胀卷积进行特征提取
  155.         # -----------------------------------------#
  156.         self.aspp = ASPP(dim_in=in_channels, dim_out=256, rate=16 // downsample_factor)
  157.  
  158.         # ----------------------------------#
  159.         #   浅层特征边
  160.         # ----------------------------------#
  161.         self.shortcut_conv = nn.Sequential(
  162.             nn.Conv2d(low_level_channels, 48, 1),
  163.             nn.BatchNorm2d(48),
  164.             nn.ReLU(inplace=True)
  165.         )
  166.  
  167.         self.cat_conv = nn.Sequential(
  168.             nn.Conv2d(48 + 256, 256, 3, stride=1, padding=1),
  169.             nn.BatchNorm2d(256),
  170.             nn.ReLU(inplace=True),
  171.             nn.Dropout(0.5),
  172.  
  173.             nn.Conv2d(256, 256, 3, stride=1, padding=1),
  174.             nn.BatchNorm2d(256),
  175.             nn.ReLU(inplace=True),
  176.  
  177.             nn.Dropout(0.1),
  178.         )
  179.         self.cls_conv = nn.Conv2d(256, num_classes, 1, stride=1)
  180.  
  181.     def forward(self, x):
  182.         H, W = x.size(2), x.size(3)
  183.         # -----------------------------------------#
  184.         #   获得两个特征层
  185.         #   low_level_features: 浅层特征-进行卷积处理
  186.         #   x : 主干部分-利用ASPP结构进行加强特征提取
  187.         # -----------------------------------------#
  188.         low_level_features, x = self.backbone(x)
  189.  
  190.         #  mobilenetV2 返回的主干特征 进行aspp 对应nets.jpg中的 encoder
  191.         #  注意返回的 主干特征是 进行到 5个层堆叠为止, 未进行后续操作
  192.         x = self.aspp(x)
  193.         #  mobilenetV2 返回的浅层特征 进行1x1的conv  对应nets.jpg中的 decoder中左侧的那个conv
  194.         low_level_features = self.shortcut_conv(low_level_features)
  195.  
  196.         # -----------------------------------------#
  197.         #   将加强特征边上采样
  198.         #   与浅层特征堆叠后利用卷积进行特征提取
  199.         #   interpolate()  插值函数， 进行上/下采样处理 , 其中的 size 代表是输出后的 shape
  200.         # -----------------------------------------#
  201.         x = F.interpolate(x, size=(low_level_features.size(2), low_level_features.size(3)), mode='bilinear',
  202.                           align_corners=True)
  203.  
  204.         # 对应nets.jpg中的 decoder中的那个Concat
  205.         # 48, 128, 128  + 256, 128, 128 -> 304, 128, 128
  206.         # 304, 128, 128 -> 256, 128, 128
  207.         x = self.cat_conv(torch.cat((x, low_level_features), dim=1))
  208.         # 256, 128, 128 -> 2, 128, 128
  209.         x = self.cls_conv(x)
  210.         # 2, 128, 128 -> 2, 512, 512
  211.         # 将分类好的特征举证 resize成原图尺寸大小 的 特征
  212.         x = F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True)
  213.         return x
  214.

 六. 源码地址

GitHub - mcuwangzaiacm/MobileV2_DeepLabV3plus_pytorch1.2: 这是一个基于MobileV2主干的DeepLabV3plus语义分割模型基础代码，用于入门学习

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/繁依Fanyi0/article/detail/678192
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号