A hybrid convolution network for serial number recognition on banknotes

As the sole identity of banknote, serial number has played a crucial role in monitoring the circulation of currencies. Serial number recognition plays an important role in financial market, which requires fast and accurate performances in real applications. In this paper, a hybrid convolution network model has been proposed, in which a dilated-based convolution neural network is employed to improve the recognition accuracy and a quantitative neural network method is developed to speed up the identification process. In dilated-based convolution neural network, the convolution layer and the pooling layer have been replaced by dilated convolution, which can reduce the computation cost. The quantitative neural network based method quantizes the weight parameters to an integer power of two, which transforms the original multiplication operation to a shift operation and can greatly reduce the time. The proposed model was examined and tested on four different banknotes with 35,000 banknote images including RMB, HKD, USD and GBP. The experimental results show that, the proposed model can efficiently improve the recognition accuracy to 99.89% and reduce the recognition time to less than 0.1 ms, and it outperforms the other algorithms on both recognition accuracy and recognition speed. (C) 2019 Elsevier Inc. All rights reserved.