Design of a Very Compact CNN Classifier for Online Handwritten Chinese Character Recognition Using DropWeight and Global Pooling
This work addresses the storage bottleneck for deploying deep learning models on portable devices, specifically for Chinese character recognition, though it is incremental as it builds on existing CNN architectures.
The paper tackled the problem of large storage requirements for CNN-based online handwritten Chinese character recognition on mobile devices by proposing DropWeight for pruning connections and using global pooling, resulting in a model requiring only 0.57 MB with a performance decrease of only 0.91% compared to state-of-the-art methods needing up to 135 MB.
Currently, owing to the ubiquity of mobile devices, online handwritten Chinese character recognition (HCCR) has become one of the suitable choice for feeding input to cell phones and tablet devices. Over the past few years, larger and deeper convolutional neural networks (CNNs) have extensively been employed for improving character recognition performance. However, its substantial storage requirement is a significant obstacle in deploying such networks into portable electronic devices. To circumvent this problem, we propose a novel technique called DropWeight for pruning redundant connections in the CNN architecture. It is revealed that the proposed method not only treats streamlined architectures such as AlexNet and VGGNet well but also exhibits remarkable performance for deep residual network and inception network. We also demonstrate that global pooling is a better choice for building very compact online HCCR systems. Experiments were performed on the ICDAR-2013 online HCCR competition dataset using our proposed network, and it is found that the proposed approach requires only 0.57 MB for storage, whereas state-of-the-art CNN-based methods require up to 135 MB; meanwhile the performance is decreased only by 0.91%.