Positive and negative max pooling for image classification

Abstract

Max pooling has been regard as the best pooling method in image classification when image features are coded by sparse coding [2]. However, max pooling reduces the classification discrimination, since it doesn't distinguish the sign of coding coefficient but only selects the max absolute value. In order to increase the image representation discrimination, we preserve the sign of code coefficient and develop a feature pooling method named PN-Max pooling. Experimental results show that PN-Max pooling achieves higher image classification accuracy than Max pooling.