Multi-Lane Capsule Network for Classifying Images With Complex Background
Siwei Chang, Jin Liu
Abstract
Capsule Network (CapsNet) is a novel structure for deep neural network, mapping the region of target instance to vectors and matrices rather than scalars. This process is enabled by dynamic routing algorithm which help CapsNet to achieve more robust capacity with fewer parameters than traditional CNNs. However, one drawback of Capsule is that it turns to account for everything in the image, which leads to a poor performance when the backgrounds are too varied to model with a reasonable sized net. We proposed a multi-lane capsule network with strict-squash (MLSCN) to solve this problem. In MLSCN, we designed a novel Capsule based network structure, replaced the Squash function and optimized the implementation of dropout. To validate modification proposed in the paper, we conducted extensive experiments on three public datasets which include MNIST, affNIST and CIFAR10. Ablation experiments were also conducted to analyze contributions of each component of MLSCN. Experimental results show that MLSCN outperforms the original CapsNet in multiple benchmarks. Our model boosts the classification accuracy of CIFAR10 about 17% with negligible parameter increase compare to original CapsNet. Besides, the proposed model achieves an accuracy of 65.37%, while original CapsNet is only 41.62% on MNIST-BC.