Chapter 9 of Huawei Deep Learning Course - Convolutional Neural Network and Case Practice

目录

前言

第一节 Convolutional Neural Networks Review Summary

第二节 经典的CNN网络

案例实践 

Resnet-经典卷积神经网络

---

前言

This course refers to learning Huawei Deep Learning Course

https://education.huaweicloud.com/courses/course-v1:HuaweiX+CBUCNXE088+Self-paced/courseware/bfde64ad93ab4bb8a0d4937bdfb29f20/f0525ab36b3e421395671cd760a4adf5/

第一节 Convolutional Neural Networks Review Summary

This section briefly describes the network layers of a convolutional neural network、The core ideas and advantages are reviewed and summarized

首先,Convolutional Neural Networks have convolutional layers、池化层和全连接层.

The function of the convolutional layer is to extract the local features of the image;

The role of the pooling layer is to reduce the amount of parameters and solve the problem of image invariance;

The role of the fully connected layer is to map the feature map to a vector,以便分类.

其次,The core idea of ​​convolutional neural network is parameter sharing and local perception.

因此,Convolutional Neural Networks imitate the characteristics of people's local perception of images,Feature extraction is performed through the action of convolution,in order to reduce the redundant amount of information,Understand the salient features of retention.

最后,Convolutional Neural Networks have significant advantages in the field of computer vision,Because it can achieve high-precision recognition and can automatically extract features（The traditional image processing is to extract features based on certain rules）.

第二节 经典的CNN网络

The development of deep learning and convolutional neural networks is inseparableILSVRC比赛,with the competition,Many classic convolutional neural networks have emerged.

AlexNet

(1)网络结构：

Because at that time the network structure appeared,Chip performance is not very powerful,所以最早AlexNetThe structure adopts the design of data shunting.

（2）特点：

• 使用ReLU函数代替了Sigmoid激活函数,实现快速收敛.
• design overlappingPooling,即stride<kernel size,Overfitting can be reduced.
• 进行了数据增强,能够防止过拟合.
• designed a stepDropout,Randomly drop certain neurons,防止过拟合. 

VGGNet 

（1）网络结构：

（2）特点：

VGGNetNeural Networks starring research on the effect of depth on convolutional networks,使用3*3卷积核,Convolution is repeated continuously、池化,最后全连接、Softmax,Get the output probability.

Below are its six configurations：

GoogLeNet

（1）网络结构：

（2）特点：

GoogLeNetWhat is explored are convolution kernels of different sizes、The impact of wider network layers on classification performance.It is proved that more convolutions and deeper layers lead to better structure. 

与AlexNetOne of the differences is joiningInception结构：

使用不同大小的卷积核,Extract features of different receptive fields.

ResNet 

（1）网络结构：

（2）特点：

The residual structure is designed,It effectively solves the problem of gradient disappearance of deep neural network.

SENet 

Not a complete neural network by itself,is a module that can be embedded into other neural networks：

Its structure is added toInceptionThe structure is as follows：

加入到残差结构中： 

 

案例实践 

Resnet-经典卷积神经网络

http://t.csdn.cn/OSiKu