Local-binarized very deep residual network for visual categorization

doi:10.1016/j.neucom.2020.11.041

	Local-binarized very deep residual network for visual categorization
	Liu, Xuejing 1,2; Li, Liang 1; Wang, Shuhui 1; Zha, Zheng-Jun 3; Huang, Qingming 1,2
刊名	NEUROCOMPUTING
	2021-03-21
卷号	430 页码:82-93
关键词	Network compression and acceleration Pose estimation Object recognition Saliency detection Local binary residual block
ISSN号	0925-2312
DOI	10.1016/j.neucom.2020.11.041
英文摘要	Residual networks usually require more layers to achieve remarkable performance in complex visual categorization tasks, such as pose estimation. However, the increasing number of layers leads to a heavy burden on training and forward inference as well as over-fitting. This paper proposed local binary residual block (LBB) to promote the very deep residual networks on the trainable parameters, FLOPs and accuracy. In each LBB, the 3 x 3 filters are binarized based on Bernoulli distribution under a sparse constraint, an activation function is prepared to trigger the non-linear response, and the linear 1 x 1 filters are learned in a real-valued way. After stochastic binarized initialization, the 3 x 3 filters in LBB need not be updated during training. The above architecture reduces at least 69.2% trainable parameters and 70.5% FLOPs compared to the original model. The LBB is derived from three observations: 1) Activated responses of one standard k x k convolutional layer can be approximated by combining binarized k x k filters with 1 x 1 filters; 2) Most computation in the very deep residual networks is spent on the 3 x 3 convolutions; and 3) 1 x 1 filters play an important role in cross-channel information integration. In addition, the LBB module is suitable for the very deep network framework, including stacked hourglass network and pyramid residual modules. Experiments are conducted on MPII and LSP dataset for pose estimation task; CIFAR-10, CIFAR-100 and ImageNet datasets for object recognition; ECSSD, HKU-IS, PASCAL-S, DUT-OMRON, DUTS for saliency detection. The results show that our model can accelerate the training and inference of the network with only a slight performance degradation. (c) 2020 Elsevier B.V. All rights reserved.
资助项目	National Key R&D Program of China[2018YFE0303104] ; National Natural Science Foundation of China[61732007] ; National Natural Science Foundation of China[61771457]
WOS研究方向	Computer Science
语种	英语
出版者	ELSEVIER
WOS记录号	WOS:000617365300008
内容类型	期刊论文
源URL	[http://119.78.100.204/handle/2XEOYT63/16230]
专题	中国科学院计算技术研究所
通讯作者	Li, Liang
作者单位	1.Chinese Acad Sci, Inst Comp Technol, CAS, Key Lab Intelligent Informat Proc, Beijing 100190, Peoples R China 2.Univ Chinese Acad Sci, Sch Comp & Control Engn, Beijing 100190, Peoples R China 3.Univ Sci & Technol China, Sch Informat Sci & Technol, Hefei 230027, Peoples R China
推荐引用方式 GB/T 7714	Liu, Xuejing,Li, Liang,Wang, Shuhui,et al. Local-binarized very deep residual network for visual categorization[J]. NEUROCOMPUTING,2021,430:82-93.
APA	Liu, Xuejing,Li, Liang,Wang, Shuhui,Zha, Zheng-Jun,&Huang, Qingming.(2021).Local-binarized very deep residual network for visual categorization.NEUROCOMPUTING,430,82-93.
MLA	Liu, Xuejing,et al."Local-binarized very deep residual network for visual categorization".NEUROCOMPUTING 430(2021):82-93.