【读书1】【2017】MATLAB与深度学习——卷积层(2)

如前一节所述，通过训练确定滤波器矩阵的值。

As addressed in the previous section, thevalues of the filter matrix are determined through the training process.

因此，这些值在整个训练过程中不断训练。

Therefore, these values are continuouslytrained throughout the training process.

这类似于普通神经网络连接权值的更新过程。

This aspect is similar to the updatingprocess of the connection weights of the ordinary neural network.

卷积是一种相当困难的运算操作，课本上的解释是因为它位于二维平面上。

The convolution is a rather difficultoperation to explain in text as it lies on the two-dimensional plane.

然而，与卷积刚出现的那个时代相比，现在，它的概念和计算步骤已经简单多了。

However, its concept and calculation stepsare simpler than they appear.

一个简单的例子将帮助你理解它是如何工作的。

A simple example will help you understandhow it works.

考虑一个4x4像素的图像，该图像表示为图6-5所示的矩阵。

Consider a 4x4 pixel image that is expressedas the matrix shown in Figure 6-5.

图6-5 4x4像素图像Four-by-four pixel image

我们将通过该图像的卷积滤波操作生成特征映射。

We will generate a feature map via theconvolution filter operation of this image.

我们使用这里所示的两个卷积滤波器。

We use the two convolution filters shownhere.

应该注意的是，实际的ConvNet滤波器是通过训练来确定的，而不是通过人工决策来确定的。

It should be noted that the filters of theactual ConvNet are determined through the training process and not by manualdecision.

让我们从第一个滤波器开始。

Let’s start with the first filter.

卷积操作从与卷积滤波器大小相同的左上角的子矩阵开始（参见图6-6）。

The convolution operation begins at theupper-left corner of the submatrix that is the same size as the convolutionfilter (see Figure 6-6).

图6-6 卷积运算从左上角开始The convolution operationstarts at the upper-left corner

卷积运算是位于两个矩阵相同位置的元素乘积之和。

The convolution operation is the sum of theproducts of the elements that are located on the same positions of the twomatrices.

图6-6中输出7的运算过程为：

The result of 7 in Figure 6-6 is calculated as:

(1x1)+(1x0)+(4x0)+(6x1) = 7

对下一个子矩阵进行卷积运算（见图6-7）。

Another convolution operation is conductedfor the next submatrix (see Figure 6-7).

图6-7 第二次卷积操作The second convolutionoperation

由设计者决定为每次操作跨越多少元素。

The designer decides how many elements tostride for each operation.

当滤波器较大时，跨越元素的个数可以大于1。

It can be greater than one if the filter islarger.

——本文译自Phil Kim所著的《Matlab Deep Learning》

更多精彩文章请关注微信号：