Like other problems in computer vision, offline handwritten Chinese characterrecognition (HCCR) has achieved impressive results using convolutional neuralnetwork (CNN)-based methods. However, larger and deeper networks are needed todeliver state-of-the-art results in this domain. Such networks intuitivelyappear to incur high computational cost, and require the storage of a largenumber of parameters, which renders them unfeasible for deployment in portabledevices. To solve this problem, we propose a Global Supervised Low-rankExpansion (GSLRE) method and an Adaptive Drop-weight (ADW) technique to solvethe problems of speed and storage capacity. We design a nine-layer CNN for HCCRconsisting of 3,755 classes, and devise an algorithm that can reduce thenetworks computational cost by nine times and compress the network to 1/18 ofthe original size of the baseline model, with only a 0.21% drop in accuracy. Intests, the proposed algorithm surpassed the best single-network performancereported thus far in the literature while requiring only 2.3 MB for storage.Furthermore, when integrated with our effective forward implementation, therecognition of an offline character image took only 9.7 ms on a CPU. Comparedwith the state-of-the-art CNN model for HCCR, our approach is approximately 30times faster, yet 10 times more cost efficient.
展开▼
