Pooling Operations in Deep Learning: From “Invariable” to “Variable”

Table 2

Summary of improved pooling operation.

Pooling method	Characteristic	Sketch map
Matrix2-norm pooling [7]	Matrix 2-norm pooling operation takes the energy of the image as the information transmitted to the next layer network to make geometric distortion of the image is highly invariable
Moment pooling [8]	The randomness of moment pooling operation makes each choice different, to prevent over inhibition
Covariance pooling [9]	Covariance pooling operation can capture more information on the feature map
Dynamic adaptive pooling [11]	Dynamic adaptive pooling operation can adaptively adjust the weight of pooling and extract more accurate features
Nested invariance pooling [12]	Nested invariance pooling operation can be extended to any arbitrary transformation set, allowing any arbitrary output feature dimension to be specified
Spectral pooling [15]	Spectral pooling operation has higher efficiency and lower cost	—