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

Table 4

Summary of variable of the pooling kernel.

Pooling method	Characteristic		Sketch map
Generalized max pooling [24]	Generalized max pooling operation balances the influence of frequent pixels and rare data and improves the ability to extract fine-grained data
Parameter pooling [25]	Parameter pooling operation converts the correlation operation into an interpretable pooling operation, which retains information and reduces errors
LP pooling [26]	LP pooling operation balances the effects of max pooling and average pooling.
Spatial attention pooling [27]	Spatial attention pooling operation mitigates the effects of distracting factors and focuses on meaningful parts of the image
Rank-based pooling [28]	Rank-based average pooling	Alleviate the problem of loss of information in max pooling and loss of discriminative information in average pooling
	Rank-based weighted pooling	Assigning a weight value to each pixel in the pooling domain can improve performance
	Rank-based stochastic pooling	Alleviates the problem that random pooling is limited to nonnegative values and reduces overfitting
Stochastic pooling [32]	Stochastic pooling operation is simple and has a strong generalization ability
Spatial pyramid pooling [33]	Spatial pyramid pooling operation can handle images of different scales, is very flexible to use, and can effectively prevent overfitting. However, there are also practical constraints to consider