收录:
摘要:
Densely connected convolutional networks (DenseNet) have reached unprecedented parameter-performance efficiencies and alleviated problems of vanishing gradients by concatenating each layer to every other layer. However, with the increase in network depth, the cross-channel interaction of dense blocks has become increasingly complex. Hence, it is now more difficult to optimize networks. Moreover, the way combining features in DenseNet restricts its flexibility and scalability in learning more expressive combination strategies. In this study, we aim to answer the question of how to simultaneously ensure the benefits of feature reuse, reduce the complexity of cross-channel interactions, and increase the flexibility of the network. Hence, the components of DenseNet are refined and then used as a basis to develop a universal densely connected convolutional network (UDenseNet). Based on the proposed architecture, the impact of different component configurations on the network performance is empirically analyzed to determine the optimal architectural configuration. Extensive experiments are conducted to validate the proposed UDenseNet on benchmark datasets (CIFAR, SVHN and ImageNet). Results show that, compared to most other methods, the proposed UDenseNet can significantly improve performance in image recognition tasks. © 2020, Springer Nature Switzerland AG.
关键词:
通讯作者信息:
电子邮件地址: