Efficient densely connected convolutional neural networks

Abstract

Listen

Translate article

Recent works have shown that convolutional neural networks (CNNs) are parameter redundant, which limits the application of CNNs in Mobile devices with limited memory and computational resources. In this paper, two novel and efficient lightweight CNNs architectures are proposed, which are called DenseDsc and Dense2Net. Two proposed CNNs are densely connected and the dense connectivity facilitates feature re-use in the networks. Dense2Net adopts efficient group convolution and DenseDsc adopts more efficient depthwise separable convolution. The novel dense blocks of DenseDsc and Dense2Net improve the parameter efficiency. The proposed DenseDsc and Dense2Net are evaluated on highly competitive classification benchmark datasets (CIFAR and ImageNet). The experimental results show that DenseDsc and Dense2Net have higher accuracy than DenseNet with similar parameters or FLOPs. Compared with other efficient CNNs with less than 0.5 M parameters for CIFAR, Dense2Net and DenseDsc achieve state-of-the-art results on CIFAR-10 and CIFAR-100, respectively. DenseDsc and Dense2Net are very competitive in efficient CNNs with less than 1.0 M parameters on CIFAR. Furthermore, Dense2Net achieves state-of-the-art results on ImageNet in manual CNNs with less than 10 M parameters.