圖6-13 當影象矩陣與濾波器不匹配時，較大的重要元素不會起到顯著的作用Whenthe image matrix does not match the filter, the significant elements are notaligned

這是因為影象矩陣與濾波器不匹配，影象矩陣的重要元素在錯誤的方向上對齊。

This is because the image matrix does notmatch the filter; the significant elements of the image matrix are aligned inthe wrong direction.

以同樣的方式處理第二個卷積濾波器，產生如圖6- 14所示的特徵對映。

In the same manner, processing the secondconvolution filter produces the feature map shown in Figure 6-14.

圖6-14 卷積輸出的大小取決於影象矩陣與卷積濾波器是否匹配The valuesdepend on whether the image matrix matches the convolution filter

與第一次卷積操作類似，該特徵對映中的元素值取決於影象矩陣是否與卷積濾波器匹配。

Similarly to the first convolutionoperation, the values in the elements of this feature map depend on whether theimage matrix matches the convolution filter or not.

綜上所述，卷積層實現輸入影象與卷積濾波器的運算併產生特徵對映。

In summary, the convolution layer operatesthe convolution filters on the input image and produces the feature maps.

在卷積層中提取的特徵由經過訓練的卷積濾波器確定。

The features that are extracted in theconvolution layer determined by the trained convolution filters.

因此，卷積層提取的特徵取決於使用哪一種卷積濾波器。

Therefore, the features that theconvolution layer extracts vary depending on which convolution filter is used.

卷積濾波器建立的特徵對映在該層產生輸出之前通過啟用函式進行處理。

The feature map that the convolution filtercreates is processed through the activation function before the layer yieldsthe output.

卷積層的啟用函式與普通神經網路的啟用函式相同。

The activation function of the convolutionlayer is identical to that of the ordinary neural network.

雖然ReLU函式在當前的大多數應用中使用，但sigmoid函式和tanh函式也經常被使用。

Although the ReLU function is used in mostof the recent applications, the sigmoid function and the tanh function areoften employed as well.

僅供參考，在數字訊號處理領域中廣泛應用的滑動平均濾波器是一種特殊的卷積濾波器。

Just for the reference, the moving averagefilter, which is widely used in the digital signal processing field, is aspecial type of convolution filter.

如果你熟悉數字濾波器，將它們與之聯絡起來，可以更好地理解卷積濾波器蘊含的思想。

If you are familiar with digital filters,relating them to this concept may allow you to better understand the ideasbehind the convolution filter.

池化層（Pooling Layer）

池化層的作用是減小影象的大小，將影象特定區域的相鄰畫素合成為單個代表值。

The pooling layer reduces the size of theimage, as it combines neighboring pixels of a certain area of the image into asingle representative value.

池化是許多其它影象處理方案中已經採用的典型技術。

Pooling is a typical technique that manyother image processing schemes have already been employing.

為了在池化層中進行操作，我們應該確定如何從影象中選擇池化畫素以及如何計算代表值。

In order to conduct the operations in thepooling layer, we should determine how to select the pooling pixels from theimage and how to set the representative value.

通常是從正方形矩陣中選擇相鄰的畫素，而畫素合成的數量因具體問題而異。

The neighboring pixels are usually selectedfrom the square matrix, and the number of pixels that are combined differs fromproblem to problem.

代表值通常被設定為所選畫素的平均值或最大值。

The representative value is usually set asthe mean or maximum of the selected pixels.

池化層的操作非常簡單。

The operation of the pooling layer issurprisingly simple.

——本文譯自Phil Kim所著的《Matlab Deep Learning》

更多精彩文章請關注微訊號：