小结(Summary)
本章涵盖以下概念:
This chapter covered the followingconcepts:
为了提高机器学习的图像识别性能,应该提供突出独特特征的特征映射,而不是原始图像。
In order to improve the image recognitionperformance of Machine Learning, the feature map, which accentuates the uniquefeatures, should be provided rather than the original image.
传统的特征提取器是人工设计的。
Traditionally, the feature extractor hadbeen manually designed.
ConvNet包含用于特征提取器的一种特殊类型的神经网络,其权重是通过训练过程确定的。
ConvNet contains a special type of neuralnetwork for the feature extractor, of which the weights are determined via thetraining process.
ConvNet由特征提取器和分类神经网络组成。
ConvNetconsists of a feature extractor and classification neural network.
它的深层架构是一个障碍,使得训练过程变得困难。
Its deep layer architecture had been abarrier that made the training process difficult.
然而,由于深度学习被引入到这个问题的解决方案中,从而导致ConvNet的使用得到迅速扩展。
However, since Deep Learning was introducedas the solution to this problem, the use of ConvNet has been growing rapidly.
ConvNet的特征提取器由卷积层和池化层的交互堆栈组成。
The feature extractor of ConvNet consistsof alternating stacks of the convolution layer and the pooling layer.
当ConvNet处理二维图像时,它的大部分操作都是在二维平面中进行的。
As ConvNet deals with two-dimensionalimages, most of its operations are conducted in a two-dimensional conceptualplane.
通过使用卷积滤波器,卷积层生成强调输入图像特征的图像映射。
Using the convolution filters, theconvolution layer generates images that accentuate the features of the inputimage.
该层的输出图像数目与网络包含的卷积滤波器的数目相同。
The number of output images from this layeris the same as the number of convolution filters that the network contains.
卷积滤波器实际上就是二维矩阵。
The convolution filter is actually nothingbut a two-dimensional matrix.
池化层减小了图像的尺寸。
The pooling layer reduces the image size.
它绑定相邻像素并用代表值替换它们。
It binds neighboring pixels and replacesthem with a representative value.
代表值一般是相邻像素的最大值或平均值。
The representative value is either themaximum or mean value of the pixels.
全书完
——本文译自Phil Kim所著的《Matlab Deep Learning》
更多精彩文章请关注微信号: