Abstract: A problem of deep convolutional neural networks is that the channel numbers of the feature maps often increases with the depth of the network. This problem can result in a dramatic increase in the number of parameters and serious over-fitting. The 1×1 convolutional layer whose kernel size is 1×1 is popular for decreasing the channel numbers of the feature maps by offer a channel-wise parametric pooling, often called a feature map pooling or a projection layer. However, the 1×1 convolutional layer has numerous parameters that need to be learned. Inspired by the 1×1 convolutional layer, we proposed a channel max pooling, which reduces the feature space by compressing multiple feature maps into one feature map via selecting the maximum values of the same locations from different feature maps. The advantages of the proposed method are twofold as follows: the first is that it decreases the channel numbers of the feature maps whilst retaining their salient features, and the second is that it non-parametric which has no increase in parameters. The experimental results on three image classification datasets show that the proposed method achieves good performance and significantly reduced the parameters of the neural network.