Pyramidal RoR for image classification

作者：Zhang, Ke^*; Guo, Liru; Gao, Ce; Zhao, Zhenbing

来源：Cluster Computing, 2019, 22(2): S5115-S5125.

DOI：10.1007/s10586-017-1443-x

• Image classification
• Residual Networks of Residual Networks
• PyramidNet
• Pyramidal RoR

The Residual Networks of Residual Networks (RoR) exhibits excellent performance in the image classification task, but sharply increasing the number of feature map channels makes the characteristic information transmission incoherent, which losses a certain of information related to classification prediction, limiting the classification performance. In this paper, a Pyramidal RoR network model is proposed by analysing the characteristics of RoR and combining with the PyramidNet. Firstly, based on RoR, the Pyramidal RoR network model with channels gradually increasing is designed. Secondly, we analysed the effect of different residual block structures on performance, and chosen the residual block structure which best favoured the classification performance. Finally, we add an important principle to further optimize Pyramidal RoR networks, drop-path is used to avoid over-fitting and save training time. In this paper, image classification experiments were performed on CIFAR-10/100, SVHN and Adience datasets, and we achieved the current lowest classification error rates were 2.96, 16.40 and 1.59% on CIFAR-10/100 and SVHN, respectively. Experiments show that the Pyramidal RoR network optimization method can improve the network performance for image classification and effectively suppress the gradient disappearance problem in DCNN training.

• 出版日期2019-3
• 单位华北电力大学