Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition. (December 2017)

• Record Type: Journal Article
• Authors: Xiao, Xuefeng; Jin, Lianwen; Yang, Yafeng; Yang, Weixin; Sun, Jun; Chang, Tianhai
• Is Part Of: Pattern recognition. Volume 72(2017:Dec.)
• Page Start: 72
• Page End: 81
• Publication Date: 2017-12
• Subjects: Convolutional neural network -- Handwritten Chinese character recognition -- CNN acceleration -- CNN compression; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.06.032
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a new method for building fast and compact CNN model for large scale handwritten Chinese character recognition (HCCR). We propose a new technique, namely Adaptive Drop-weight (ADW), for effectively pruning CNN parameters. We proposed the Global Supervised Low Rank Expansions (GSLRE) method for accelerating CNN model. Comparing with the state-of-the-art CNN method for HCCR, our approach is about 30-times faster yet 10-times smaller. Abstract: Like other problems in computer vision, offline handwritten Chinese character recognition (HCCR) has achieved impressive results using convolutional neural network (CNN)-based methods. However, larger and deeper networks are needed to deliver state-of-the-art results in this domain. Such networks intuitively appear to incur high computational cost, and require the storage of a large number of parameters, which render them unfeasible for deployment in portable devices. To solve this problem, we propose a Global Supervised Low-rank Expansion (GSLRE) method and an Adaptive Drop-weight (ADW) technique to solve the problems of speed and storage capacity. We design a nine-layer CNN for HCCR consisting of 3755 classes, and devise an algorithm that can reduce the network's computational cost by nine times and compress the network to 1/18 of the original size of the baseline model, with only a 0.21% drop in accuracy. In tests, the proposed algorithm can still surpass the best single-network performance reported thus far in the literature while requiring only 2.3MB for storage. Furthermore, when integrated with our effective forward implementation, the recognition of an offline character image takes only 9.7 ms on a CPU. Compared with the state-of-the-art CNN model for HCCR, our approach is approximately 30 times faster, yet 10 times more cost efficient.