Convolutional neural network (Introduction)

One 、 A brief introduction of convolutional neural network

adopt Input layer Input , Then according to several Convolution layer Just repeat the following steps several times , Last Output layer Output .

A basic process in convolution calculation is ： Convolution -->ReLU（ Modified linear element ）--> Pooling （ Down sampling ）

Then proceed Full connection , Calculate the probability according to the weight , Judge .

Two 、“ Detailed explanation ”

1. Convolution calculation

Split the image into corresponding feature points , go by the name of Convolution kernel .

Then check whether the recognized image has a corresponding convolution kernel to confirm whether it is the target object .

The target image passes Convolution kernel A two-dimensional figure obtained is Characteristics of figure .

2. Activation function

frequently-used ：Relu：if x<0: y=0 else y=x

Convert all the negative eigenvalues calculated by convolution into 0, Will not change the characteristics .

3. Pooling

There are a lot of details about one （ It's a feature ）, Or a considerable number of hierarchical details , The complexity of the algorithm is very high .

So there is Pooling ( layer )（pooling）. Pooling is to reduce the characteristic matrix , namely Zoom out feature map （Feature Map）

Two pooling methods ：

1. Maximum pooling ： Select the maximum value in the scanned area as Feature Map A characteristic value of
2. The average pooling ： Take the average value in the scanned area as Feature Map A characteristic value of  

stay Dealing with edges The operation of is called （Padding）

If you use maximum pooling for images , be Fill zero on the edge To extract edge features

Pooling requires that the features of the original feature map must be retained

4. Full link

Operate the array according to the array weight of the target graph to obtain a probability number to judge whether it is the target . Using a large amount of data training for machine learning to correct the convolution kernel and full link behavior . Then use back propagation （backpropagation） The algorithm is constantly modified to deal with the full connection of feature arrays , Finally, we get more and more accurate network .

The convolution kernel and full connection at the beginning are random , Artificially designated , As long as enough data and feedback are fed to the network , Finally, we can get a better algorithm network .

Common English explanations in Convolutional Neural Networks

Filtering: The math behind the match Filter ： The math behind comparison

1. Line up the feature and the image patch. Feature and image complement alignment .

2. Multiply each image pixel by the corresponding feature pixel. Multiply each image pixel by the corresponding feature pixel .

3. Add them up. Add up

4. Divide by the total number of pixels in the feature. Divide by the total number of pixels in the feature （ Here refers to the number of pixels in the fragment ）.---> Get the final pixel value

Pooling: Shrinking the image stack： Pooling : Shrink the image stack  

1. Pick a window size (usually 2 or 3). Choose one 2X2 Of window

2.Pick a stride (usually 2) Move the window one step to the right 2

3.Walk your window across your filtered images. Slide window to filter pictures

4. From each window, take the maximum value. Record the maximum value in turn

After this step , We get a similar , But the smaller picture . It can still be seen that the characteristics of the maximum composition are .

The above is the summary after watching the video .

Push video （ Too cattle ！）： Explain the working principle of convolutional neural network in vernacular _ Bili, Bili _bilibili