Journal of Northeastern University Natural Science ›› 2020, Vol. 41 ›› Issue (1): 12-16.DOI: 10.12068/j.issn.1005-3026.2020.01.003

• Information & Control • Previous Articles     Next Articles

GMSDenseNet:Lightweight DenseNet Based on Group Multi-structure Convolution

YU Chang-yong, HE Xin, QI Xin, MA Hai-tao   

  1. School of Computer and Communication Engineering, Northeastern University at Qinhuangdao, Qinhuangdao 066004, China.
  • Received:2019-03-10 Revised:2019-03-10 Online:2020-01-15 Published:2020-02-01
  • Contact: HE Xin
  • About author:-
  • Supported by:
    -

Abstract: A simple and efficient lightweight DenseNet model was proposed, which optimized the parameter redundancy and high FLOPs(floating point operations) in the original DenseNet model. Firstly, the details and characteristics of existing popular convolutional computing units were analyzed. Secondly, the convolutional unit combination with good characteristics was applied to design the multi-structure convolution unit which was used to construct the lightweight DenseNet model. The complexity of this lightweight model and the original DenseNet model was further analyzed. The optimal configuration of the constructed network structure was given according to the experimental results. The results of using the DenseNet-40 model with about 18.8% of FLOPs and 28.4% of the model parameters are obtained, and the accuracy is only reduced by 0.4% at most.

Key words: group convolution, depth separable convolution, group multi-structure convolution unit, lightweight, DenseNet

CLC Number: