Learning Efficient Residual Networks through Dense Connecting
Tianzhong Song, Yan Song, Yongxiong Wang, Xuegang Huang · 2018
In this paper, a novel method is developed to study the efficient residual networks (ResNets) through the dense connecting. More precisely, in order to enhance the feature mapping as well as reducing the computing resource requirements, the dense connecting is introduced into each residual block, which then gives rise to the resulting network structures dense residual networks (DRNs). In the proposed DRNs, the feature maps of each block can be adequately used by the underlying ResNets in virtue of the employed dense connecting. Moreover, the number of filters of each convolution is greatly reduced to the half, and this practically promotes the efficiency of the model. Under the application of dense connecting, the higher performance of ResNets can be guaranteed as well as the reduction of the time-expense. Meanwhile, the addressed model is less prone to overfitting. In the end, a series of experiments on two benchmark datasets are utilized to illustrate the effectiveness of the proposed DRNs.