Depth-Wise Pooling: A Parameter-Less Solution for Channel Reduction of Feature-Map in Convolutional Neural Network
Abid Hussain, Wang Hesheng · 2019
The Convolutional neural networks (CNN) almost always consists of spatial pooling, which reduce the spatial resolution of feature-maps without any trainable parameters. This not only facilitates the network to go deeper but also reduces computational complexity. But for feature-map channels reduction the cheapest method is 1×1 convolution used before the expensive 3×3 and 5×5 convolutions. We proposed depth-wise pooling a parameter less solution for feature-map channel reduction. Which can control number of feature-map in deep CNNs. The information lost can be control with feature sharing between consecutive selected channels of feature-maps. In this work we implement four depth-wise pooling methods for reduction of feature-map channels by changing the stride size and filter size. For fair analysis we implement our methods in DenseNet [1] which has state-of-the-art results and minimum number of parameters. We showed that parameters can be reduced up to 4 times in some cases and having almost the same results.