Everything about pooling in cnn

What is max pooling?

Suppose we have a 4x4 image and we applied a 2x2 filter then we will get a 3x3 image according to convolution. So applying max-pooling means applying another filter on our output that we get after applying convolution. The reason for doing that is the filter we applied in convolution can detect the shape but it can not be detect properly or clearly but if we use max-pooling we can detect the shape properly or clearly because we are taking a high pixel value.

Suppose we get our output after convolution is:

Now we will apply our 2x2 filter to our output. We will fit our max-pooling filter first-row and first-column on the first-row and first-column of image layer, what we get after applying convolution and other cells of filter will fit in order. Then we will find the max value of present inside the filter. After getting the max value we will take it as our new output value and will put the value in the new matrix first-column first-row.

In convolution, after completing the calculation, we move the one step right but here we will move it according to the max-pooling filter size. This means if our filter is 2x2 then we will move it 2 steps right and if our filter is 3x3 then we will move it 3 steps right.

After moving the filter we will again find the max value. Our work is to find the max value and take it as output and put it on our new image layer matrix. After completing the left side we will move to one step down and fit the filter to the left side. It means now our filter should be on the image layer second-row first-column and again have to do the same work. So this way max-pooling works.

Output after applying max pooling:

What is min pooling?

The working procedure is same as max-pooling but the difference is in max-pooling we take max value. This means high pixel and in min-pooling, we take minimum value means low pixels.

What is average pooling?

The working procedure is the same as max-pooling but the difference is in max-pooling we take max value means high pixel and in average-pooling, we take average value means average pixels.

How to find average:
If we have a 2x2 filter and if we apply it on the image layer then we will see 4 values are present in the four-cell of the filter. So hare we will do average of those values to get the average value.

What is sum pooling?

The working procedure is the same as max-pooling but the difference is in max-pooling we take max value means high pixel and in sum pooling, we take sum of values present in the polling filter.

How to find sum:
If we have a 2x2 filter and if we apply it on the image layer then we will see 4 values are present in the four-cells of the filter. So hare we have do those values sum to get the sum value.