【读书1】【2017】MATLAB与深度学习——卷积层(3)

以同样的方式进行第三次卷积运算，如图6-8所示。

In the same manner, the third convolutionoperation is conducted, as shown in Figure 6-8.

图6-8 第三次卷积运算The third convolutionoperation

一旦顶部行的运算完成，从下一行左边继续开始（参见图6-9）。

Once the top row is finished, the next rowstarts over from the left (see Figure 6-9).

图6-9 卷积运算重新从左边开始The convolution operationstarts over from the left

它重复相同的过程，直到生成给定滤波器的特征映射，如图6-10所示。

It repeats the same process until thefeature map of the given filter is produced, as shown in Figure 6-10.

图6-10 给定滤波器的特征映射计算完成The feature map of thegiven filter has been completed

现在，仔细看看这个特征映射。

Now, take a closer look at the feature map.

映射中第(3, 1)位置的元素为最大值。

The element of (3, 1) of the map shows thegreatest value.

这个位置的元素究竟发生了什么呢？

What happened to this cell?

第(3, 1)位置的元素由图6-11中的卷积运算完成。

This value is the result of the convolutionoperation shown in Figure 6-11.

图6-11 图像子矩阵与卷积滤波器相匹配的运算The submatrix ofthe image matches the convolution filter

从图中可以看出，图像的子矩阵与卷积滤波器的形式相匹配，这两个矩阵都是在相同位置上具有显著数值的对角矩阵。

It is noticeable from the figure that thesubmatrix of the image matches the convolution filter; both are diagonalmatrices with significant numbers on the same cells.

当输入数据与滤波器相匹配时，卷积运算产生较大的输出值，如图6-12所示。

The convolution operation yields largevalues when the input matches the filter, as shown in Figure 6-12.

图6-12当输入数据与滤波器相匹配时，卷积运算产生较大的输出值Theconvolution operation yields large values when the input matches the filter

相比之下，在图6-13所示情况下，同样的重要数值30并没有对卷积结果产生非常大的影响，卷积输出值只有4。

In contrast, in the case shown in Figure6-13, the same significant number of 30 does not affect the convolution result,which is only 4.

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

更多精彩文章请关注微信号：