Recognition of Plate Identification Numbers Using Convolution Neural Network and Character Distribution Rules

Abstract

Recognition of plate identification numbers (PINs) is of much importance for the automation of the iron and steel production. The recognition of PINs in industrial site is a challenging problem due to complicated background, low quality of characters. Conventional image processing algorithms are employed to extract the numbers, but it is difficult for these methods to locate and recognize the numbers on the plates in complex industrial production by manually designed features. The end-to-end convolution neural network is employed to solve these problems by automatically extracted features. These features seldom combine the real production rules. A delicate recognition method of PINs using convolution neural network and characters distribution rules is proposed. The PINs are roughly recognized by convolution neural network with Non-Maximum Suppression (NMS), and the PINs are exactly processed using the character distribution rules. Experiment results demonstrate that the method proposed can arrive at a very high recognition rate 96.32% and improve the recognition rate by 54.07% compared with the end-to-end convolution neural network.