Asymmetric convolution with densely connected networks

Abstract

Convolutional neural networks are vital to some computer vision tasks, and the densely connected network is a creative architecture among them. In densely connected network, most convolution layer tends to have a much larger number of input channels than output channels, making itself to a funnel shape. We replace the 3 × 3 convolution in the densely connected network with two continuous asymmetric convolutions to make the DenseNet family more diverse. We also proposed a model in which two continuous asymmetric convolutions each outputs half of the output channels and concatenate them as the final output of these layers. Compared with the original densely connected network, our models achieve similar performance on CIFAR-10/100 dataset with fewer parameters and less computational cost.