Design of a Very Compact CNN Classifier for Online Handwritten Chinese Character Recognition Using DropWeight and Global Pooling

摘要

Currently, owing to the ubiquity of mobile devices, online handwrittenChinese character recognition (HCCR) has become one of the suitable choice forfeeding input to cell phones and tablet devices. Over the past few years,larger and deeper convolutional neural networks (CNNs) have extensively beenemployed for improving character recognition performance. However, itssubstantial storage requirement is a significant obstacle in deploying suchnetworks into portable electronic devices. To circumvent this problem, wepropose a novel technique called DropWeight for pruning redundant connectionsin the CNN architecture. It is revealed that the proposed method not onlytreats streamlined architectures such as AlexNet and VGGNet well but alsoexhibits remarkable performance for deep residual network and inceptionnetwork. We also demonstrate that global pooling is a better choice forbuilding very compact online HCCR systems. Experiments were performed on theICDAR-2013 online HCCR competition dataset using our proposed network, and itis found that the proposed approach requires only 0.57 MB for storage, whereasstate-of-the-art CNN-based methods require up to 135 MB; meanwhile theperformance is decreased only by 0.91%.