11-图像分割之RedNet

最新推荐文章于 2023-11-27 17:05:03 发布

冰冻之瓜非一日之寒

最新推荐文章于 2023-11-27 17:05:03 发布

阅读量2.2k

点赞数

分类专栏： # 从零开始图像分割文章标签：图像处理人工智能

本文链接：https://blog.csdn.net/qq_40073354/article/details/121006669

版权

从零开始图像分割专栏收录该内容

14 篇文章 19 订阅

订阅专栏

1.前置

（1）像素深度

像素深度指存储每个像素用的位数，像素深度越深，所占用存储空间越大；像素深度越浅，图像越粗糙和不自然。

（假设一幅彩色图像的每个像素用R，G，B三个分量表示，若每个分量用8位，那么一个像素共用24位表示，就说像素的深度为24）

（2）图像深度

图像深度<像素深度，接着用上面的例子，实际存储中一个像素其实是大于24位的（假设需要26位来存储）,这多出来的两位可能包含了图像的一些属性。图像深度是指实际用于存储图像的灰度或色彩所需要的比特位数（24位），而此时像素深度为26位。

（3）灰度图

灰度使用黑色调来表示物体，每个灰度对象都有从0%（白色）到100%（黑色）的亮度值，灰度图就是用灰度表示的图像（说人话灰度图就是由很多个亮度不同的黑白块组成），灰度级越多，图像层次越逼真

（4）深度图

深度图是代表距离的图像，像素值代表图像采集器到场景中各点的距离（深度）

目前深度图的研究主要有深度估计和深度图补全，深度估计以软件为基础，通过单目或双目采集到的图像转化为深度图；深度图补全以硬件为基础，通过硬件直接拍摄出的图像存在一定的误差，需要进行一定的矫正。

2.摘要

背景介绍:室内语义分割一直是计算机视觉领域的一个难题
算法组成:在RedNet中，残差模块作为基本的积木块应用于编码器和解码器，构造了一种融合结构，提出了一种金字塔监督训练方案
模型评估:实验结果表明，所提出的RedNet (Resnet-50)在SUN RGBD数据集上达到了47.8%
的最新mloU

3.Pytorch实现RedNet

from torch import nn

def conv3x3(in_planes,out_planes,stride=1):
    return nn.Conv2d(in_planes,out_planes,kernel_size=3,stride=stride,padding=1,bias=False)

class Bottleneck(nn.Module):
    expansion=4

    def __init__(self,inplanes,planes,stride=1,downsample=None):
        super(Bottleneck,self).__init__()
        self.conv1=nn.Conv2d(inplanes,planes,kernel_size=1,bias=False)
        self.bn1=nn.BatchNorm2d(planes)
        self.conv2=nn.Conv2d(planes,planes,kernel_size=3,stride=stride,padding=1,bias=False)
        self.bn2=nn.BatchNorm2d(planes)
        self.conv3=nn.Conv2d(planes,planes*4,kernel_size=1,bias=False)
        self.bn3=nn.BatchNorm2d(planes*4)
        self.relu=nn.ReLU(inplace=True)
        self.downsample=downsample
        self.stride=stride

    def forward(self,x):
        residual=x

        out=self.conv1(x)
        out=self.bn1(out)
        out=self.relu(out)

        out=self.conv2(out)
        out=self.bn2(out)
        out=self.relu(out)

        out=self.conv3(out)
        out=self.bn3(out)

        if self.downsample is not None:
            residual=self.downsample(x)

        out+=residual
        out=self.relu(out)

        return out

class TransBasicBlock(nn.Module):
    def __init__(self,inplanes,planes,stride=1,upsample=None):
        super(TransBasicBlock,self).__init__()
        self.conv1=conv3x3(inplanes,inplanes)
        self.bn1=nn.BatchNorm2d(inplanes)
        self.relu=nn.ReLU(inplace=True)
        if upsample is not None and stride!=1:
            self.conv2=nn.ConvTranspose2d(inplanes,planes,kernel_size=3,stride=stride,padding=1,output_padding=1,bias=False)
        else:
            self.conv2=conv3x3(inplanes,planes,stride)
        self.bn2=nn.BatchNorm2d(planes)
        self.upsample=upsample
        self.stride=stride

    def forward(self,x):
        residual=x

        out=self.conv1(x)
        out=self.bn1(out)
        out=self.relu(out)

        out=self.conv2(out)
        out=self.bn2(out)

        if self.upsample is not None:
            residual=self.upsample(x)

        out+=residual
        out=self.relu(out)

        return out

class RedNet(nn.Module):
    def __init__(self, num_classes=41):

        super(RedNet, self).__init__()
        block = Bottleneck
        transblock = TransBasicBlock
        layers = [3, 4, 6, 3]
        # original resnet
        self.inplanes = 64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)

        # resnet for depth channel
        self.inplanes = 64
        self.conv1_d = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3,
                                 bias=False)
        self.bn1_d = nn.BatchNorm2d(64)
        self.layer1_d = self._make_layer(block, 64, layers[0])
        self.layer2_d = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3_d = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4_d = self._make_layer(block, 512, layers[3], stride=2)

        self.inplanes = 512
        self.deconv1 = self._make_transpose(transblock, 256, 6, stride=2)
        self.deconv2 = self._make_transpose(transblock, 128, 4, stride=2)
        self.deconv3 = self._make_transpose(transblock, 64, 3, stride=2)
        self.deconv4 = self._make_transpose(transblock, 64, 3, stride=2)

        self.agant0 = self._make_agant_layer(64, 64)
        self.agant1 = self._make_agant_layer(64 * 4, 64)
        self.agant2 = self._make_agant_layer(128 * 4, 128)
        self.agant3 = self._make_agant_layer(256 * 4, 256)
        self.agant4 = self._make_agant_layer(512 * 4, 512)

        # final block
        self.inplanes = 64
        self.final_conv = self._make_transpose(transblock, 64, 3)

        self.final_deconv = nn.ConvTranspose2d(self.inplanes, num_classes, kernel_size=2,
                                               stride=2, padding=0, bias=True)

        self.out5_conv = nn.Conv2d(256, num_classes, kernel_size=1, stride=1, bias=True)
        self.out4_conv = nn.Conv2d(128, num_classes, kernel_size=1, stride=1, bias=True)
        self.out3_conv = nn.Conv2d(64, num_classes, kernel_size=1, stride=1, bias=True)
        self.out2_conv = nn.Conv2d(64, num_classes, kernel_size=1, stride=1, bias=True)

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []

        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def _make_transpose(self, block, planes, blocks, stride=1):

        upsample = None
        if stride != 1:
            upsample = nn.Sequential(
                nn.ConvTranspose2d(self.inplanes, planes,
                                   kernel_size=2, stride=stride,
                                   padding=0, bias=False),
                nn.BatchNorm2d(planes),
            )

        layers = []

        for i in range(1, blocks):
            layers.append(block(self.inplanes, self.inplanes))

        layers.append(block(self.inplanes, planes, stride, upsample))
        self.inplanes = planes

        return nn.Sequential(*layers)

    def _make_agant_layer(self, inplanes, planes):

        layers = nn.Sequential(
            nn.Conv2d(inplanes, planes, kernel_size=1,
                      stride=1, padding=0, bias=False),
            nn.BatchNorm2d(planes),
            nn.ReLU(inplace=True)
        )
        return layers

    def forward_downsample(self, rgb, depth):
        x = self.conv1(rgb)
        x = self.bn1(x)
        x = self.relu(x)

        depth = self.conv1_d(depth)
        depth = self.bn1_d(depth)
        depth = self.relu(depth)

        fuse0 = x + depth

        x = self.maxpool(fuse0)
        depth = self.maxpool(depth)

        x = self.layer1(x)
        depth = self.layer1_d(depth)

        fuse1 = x + depth

        x = self.layer2(fuse1)
        depth = self.layer2_d(depth)

        fuse2 = x + depth

        x = self.layer3(fuse2)
        depth = self.layer3_d(depth)

        fuse3 = x + depth

        x = self.layer4(fuse3)
        depth = self.layer4_d(depth)

        fuse4 = x + depth

        return fuse0, fuse1, fuse2, fuse3, fuse4

    def forward_upsample(self, fuse0, fuse1, fuse2, fuse3, fuse4):
        agent4 = self.agant4(fuse4)
        x = self.deconv1(agent4)
        if self.training:
            out5 = self.out5_conv(x)

        x = x + self.agant3(fuse3)

        x = self.deconv2(x)
        if self.training:
            out4 = self.out4_conv(x)

        x = x + self.agant2(fuse2)
        x = self.deconv3(x)
        if self.training:
            out3 = self.out3_conv(x)

        x = x + self.agant1(fuse1)
        x = self.deconv4(x)
        if self.training:
            out2 = self.out2_conv(x)

        x = x + self.agant0(fuse0)

        x = self.final_conv(x)
        out = self.final_deconv(x)

        if self.training:
            return out, out2, out3, out4, out5

        return out

    def forward(self, rgb, depth):
        fuse = self.forward_downsample(rgb, depth)
        out = self.forward_upsample(*fuse)

        return out
if __name__ == "__main__":
    import torch as t

    rgb = t.randn(1, 3, 352, 480)
    depth = t.randn(1, 1, 352, 480)

    net = RedNet()

    out = net(rgb, depth)

    print(out[0].shape)
    print(out[1].shape)
    print(out[2].shape)
    print(out[3].shape)
    print(out[4].shape)

参考：

B站深度之眼

冰冻之瓜非一日之寒

关注

0
点赞
踩
2

收藏

觉得还不错? 一键收藏
1
评论
11-图像分割之RedNet

1.前置（1）像素深度像素深度指存储每个像素用的位数，像素深度越深，所占用存储空间越大；像素深度越浅，图像越粗糙和不自然。（假设一幅彩色图像的每个像素用R，G，B三个分量表示，若每个分量用8位，那么一个像素共用24位表示，就说像素的深度为24）（2）图像深度图像深度<像素深度，接着用上面的例子，实际存储中一个像素其实是大于24位的（假设需要26位来存储）,这多出来的两位可能包含了图像的一些属性。图像深度是指实际用于存储图像的灰度或色彩所需要的比特位数（24位），而此时像素深度...
复制链接

扫一扫

专栏目录