LAUSR

Abstract Convolutional neural networks (CNNs) have attracted considerable attention in many application fields for their great ability to deal with image recognition and object detection tasks. A pooling process is an important process in CNNs, which serves to decrease the dimensionality of processed data for reducing computational cost as well as for enhancing tolerance to translation and noise. Although standard pooling methods, such as the max pooling and the average pooling, are typically adopted in many studies, a newly devised pooling method could improve the generalization ability of CNNs. In this study, we propose a hybrid pooling method which stochastically chooses the max pooling or the average pooling in each pooling layer. A characteristic of the hybrid pooling is that the probability for choosing one of the two pooling methods can be controlled for each convolutional layer. In image classification tasks with benchmark datasets, we show that the hybrid pooling is effective for increasing the generalization ability of CNNs. Moreover, we demonstrate that the hybrid pooling combined with the dropout is competitive with other existing methods in classification performance.