ConvNet的特征提取神经网络越深(包含的层数越多)，其识别效果就越好，但在训练过程中会造成一些困难，使得ConvNet在当时的技术条件下并不可行，因此一度被大家遗忘。
ConvNet yields better image recognitionwhen its feature extraction neural network is deeper (contains more layers), atthe cost of difficulties in the training process, which had driven ConvNet tobe impractical and forgotten for a while.
让我们再深入一些。
Let’s go a bit deeper.
ConvNet由提取输入图像特征的神经网络和分类特征图像的神经网络组成。
ConvNet consists of a neural network thatextracts features of the input image and another neural network that classifiesthe feature image.
图6-3给出了ConvNet的典型结构图。
Figure 6-3 shows the typical architectureof ConvNet.
6-3 ConvNet典型结构图Typical architecture of ConvNet
将输入图像数据送入特征提取网络。
The input image enters into the featureextraction network.
提取出来的特征信号进入分类神经网络。
The extracted feature signals enter theclassification neural network.
分类神经网络根据图像特征进行运算操作并产生输出。
The classification neural network thenoperates based on the features of the image and generates the output.
第4章讨论的分类技术将应用于此。
The classification techniques discussed inChapter 4 apply here.
特征提取神经网络由卷积层堆叠和池化层对组成。
The feature extraction neural networkconsists of piles of the convolutional layer and pooling layer pairs.
卷积层，顾名思义，使用卷积运算转换图像。
The convolution layer, as its name implies,converts the image using the convolution operation.
它可以被认为是数字滤波器的集合。
It can be thought of as a collection ofdigital filters.
池化层将邻近像素组合成单个像素。
The pooling layer combines the neighboringpixels into a single pixel.
因此，池化层减小了图像的尺寸。
Therefore, the pooling layer reduces thedimension of the image.
由于ConvNet的主要关注点是图像，因此卷积层和池化层的操作在概念上是在二维平面中进行的。
As the primary concern of ConvNet is theimage; the operations of the convolution and pooling layers are conceptually ina two-dimensional plane.
这是ConvNet与其它神经网络之间的区别之一。
This is one of the differences betweenConvNet and other neural networks.
综上所述，ConvNet由特征提取网络和分类网络串行连接组成。
In summary, ConvNet consists of the serialconnection of the feature extraction network and the classification network.
通过训练过程确定各层的权重。
Through the training process, the weightsof both layers are determined.
特征提取层具有成对的卷积层堆叠和池化层。
The feature extraction layer has piledpairs of the convolution and pooling layers.
卷积层通过卷积运算转换图像，并通过池化层减小图像的维数。
The convolution layer converts the imagesvia the convolution operation, and the pooling layer reduces the dimension ofthe image.
分类网络通常采用普通的多级分类神经网络。
The classification network usually employsthe ordinary multiclass classification neural network.
——本文译自Phil Kim所著的《Matlab Deep Learning》
更多精彩文章请关注微信号：