因此，在第一次接触这个概念时，即使是神经网络专家也常常难以理解。
Therefore, even the experts of neuralnetworks often have a hard time understanding this concept on their firstencounter.
也就是说，在概念和操作上，ConvNet与以前的神经网络究竟有多少不同。
That is how much ConvNet differs in conceptand operation from the previous neural networks.
本节简要介绍了ConvNet的基本结构。
This section briefly introduces thefundamental architecture of ConvNet.
从根本上来说，图像识别就是一种分类。
Basically, image recognition is theclassification.
例如，识别图片中的图像是猫还是狗，与将图像分类为猫或狗类是一样的。
For example, recognizing whether the imageof a picture is a cat or a dog is the same as classifying the image into a cator dog class.
同样的情况也适用于字母识别；从图像中识别字母，与将图像分类为字母类别中的某一类是相同。
The same thing applies to the letterrecognition; recognizing the letter from an image is the same as classifyingthe image into one of the letter classes.
因此，ConvNet的输出层一般采用多层分类的神经网络。
Therefore, the output layer of the ConvNetgenerally employs the multiclass classification neural network.
然而，无论识别方法如何，直接使用原始图像进行图像识别会导致效果不佳；应该对图像进行处理以对比它们之间的特征。
However, directly using the original imagesfor image recognition leads to poor results, regardless of the recognitionmethod; the images should be processed to contrast the features.
第4章中的例子使用了原始图像，但因为它们是简单的黑白图像，所以工作的效果很好。
The examples in Chapter 4 used the originalimages and they worked well because they were simple black-and-white images.
否则，识别结果将会很差。
Otherwise, the recognition process wouldhave ended up with very poor results.
为此，已经开发了用于图像特征提取的各种技术。
For this reason, various techniques forimage feature extraction have been developed.
代表性的方法包括SIFT、HoG、Textons、Spin image、RIFT和GLOH。
The representative methods include SIFT,HoG, Textons, Spin image, RIFT, and GLOH.
在ConvNet之前的研究中，特征提取器是由特定领域的专家设计的。
Before ConvNet, the feature extractor hasbeen designed by experts of specific areas.
因此，它需要花费大量的成本和时间，同时产生不一致的性能水平。
Therefore, it required a significant amountof cost and time while it yielded an inconsistent level of performance.
这些特征提取器独立于机器学习。
These feature extractors were independentof Machine Learning.
图6-1描述了这一实现过程。
Figure 6-1 illustrates this process.
图6-1 独立于机器学习的特征抽取器Feature extractors usedto be independent of Machine Learning
ConvNet在训练过程中包括特征提取器，而不是由专家手动设计。
ConvNetincludes the feature extractor in the training process rather than designing itmanually.
ConvNet的特征抽取器由特殊类型的神经网络组成，通过训练过程确定神经网络的权值。
The feature extractor of ConvNet iscomposed of special kinds of neural networks, of which the weights aredetermined via the training process.
ConvNet将手动特征提取转化为自动提取是其主要特征和优势。
The fact that ConvNet turned the manualfeature extraction design into the automated process is its primary feature andadvantage.
图6-2描述了ConvNet的训练过程。
Figure 6-2 depicts the training concept ofConvNet.
图6-2 ConvNet特征提取器由特殊类型的神经网络构成ConvNet’s featureextractor is composed of special kinds of neural networks
——本文译自Phil Kim所著的《Matlab Deep Learning》
更多精彩文章请关注微信号：