Down sampling and up sampling

1、 What is the function of down sampling ？ What are the usual ways ？

The lower sampling layer has two functions , One is to reduce the amount of calculation , Prevent over fitting ; The second is to increase the receptive field , So that the later convolution kernel can learn more global information .

There are two main methods of down sampling ：
1、 use stride by 2 Pooled layer , Such as Max-pooling and Average-pooling, At present, we usually use Max-pooling, Because it is easy to calculate and can better preserve texture features ;
2、 use stride by 2 The convolution of layer , The process of down sampling is a process of information loss , The pool layer is not learnable , use stride by 2 Instead of learning convolution layer pooling Can get better effect , Of course, it also increases the amount of calculation .

2、 Principle and common methods of up sampling
 

In a convolutional neural network , Because the input image passes through the convolution neural network (CNN) After feature extraction , The size of the output tends to get smaller , Sometimes we need to restore the image to its original size for further calculation ( Such as image semantic segmentation ), The operation of mapping an image from a small resolution to a large resolution , It's called upsampling , There are generally three ways to implement it ：

• interpolation , Bilinear interpolation is generally used , Because the effect is the best , Although computationally more complex than other interpolation methods , But compared with convolution calculation, it is not worth mentioning , Other interpolation methods include nearest neighbor interpolation 、 Trilinear interpolation, etc ;

• Transpose convolution or deconvolution (Transpose Conv), By inputting feature map Interval filling 0, Then carry out standard convolution calculation , Can make the output feature map The size of is larger than the input ;

• Up-Pooling - Max Unpooling && Avg Unpooling --Max Unpooling, In symmetrical max pooling The index position of the maximum value of the position record , And then in unpooling During the phase, the corresponding value is placed at the original maximum position , Make up for the rest 0;