Searching for just a few words should be enough to get started. If you need to make more complex queries, use the tips below to guide you.
Article type: Research Article
Authors: Guo, Zhendong | Li, Xiaohong | Zhang, Kai | Guo, Xiaoyong; *
Affiliations: Tianjin University of Science and Technology, College of Electronic Information and Automation, Tianjin, China
Correspondence: [*] Corresponding author. Xiaoyong Guo, Tianjin University of Science and Technology, College of Electronic Information and Automation, 1038 Dagu Nanlu, Hexi District, Tianjin, China, 300457. E-mail: gxyauthor@tust.edu.cn.
Abstract: In this paper, it is proposed that the redundancy in convolutional neural networks of object detection can be effectively removed via an adaptive pruning threshold method (APTCNN) which is associated with scaling factors in batch normalization layers. In this way, the channel pruning can be done iteratively with varying pruning threshold until the satisfactory performance is obtained. The method is also useful for identifying the unimportant convolutional layers. Therefore it can be applied for layer pruning. The experiments are conducted on three benchmark object detection datasets. APTCNN is verified for pruning the backbone network of object detectors YOLOv3 and YOLOv3-spp. It is shown that the importance of channels and layers are accurately ranked by the proposed adaptive threshold. For the channel pruning, our method reduces the size of YOLOv3 and YOLOv3-spp by 32× and 48× respectively, and accelerates 1.7× and 1.9× respectively. However, the accuracy suffers only 0.77% and 1.32% loss, respectively. As a result, the redundancy in the network architecture can be efficiently removed yielding a slimmed model that has lower computing operations, reduced size, and without compromising accuracy.
Keywords: Object detection, adaptive pruning threshold, channel pruning, layer pruning
DOI: 10.3233/JIFS-213002
Journal: Journal of Intelligent & Fuzzy Systems, vol. 43, no. 6, pp. 7821-7831, 2022
IOS Press, Inc.
6751 Tepper Drive
Clifton, VA 20124
USA
Tel: +1 703 830 6300
Fax: +1 703 830 2300
sales@iospress.com
For editorial issues, like the status of your submitted paper or proposals, write to editorial@iospress.nl
IOS Press
Nieuwe Hemweg 6B
1013 BG Amsterdam
The Netherlands
Tel: +31 20 688 3355
Fax: +31 20 687 0091
info@iospress.nl
For editorial issues, permissions, book requests, submissions and proceedings, contact the Amsterdam office info@iospress.nl
Inspirees International (China Office)
Ciyunsi Beili 207(CapitaLand), Bld 1, 7-901
100025, Beijing
China
Free service line: 400 661 8717
Fax: +86 10 8446 7947
china@iospress.cn
For editorial issues, like the status of your submitted paper or proposals, write to editorial@iospress.nl
如果您在出版方面需要帮助或有任何建, 件至: editorial@iospress.nl