A New Approach for Character Recognition of Multi-Style Vehicle License Plates

Abstract

The recognition of vehicle license plate is an important part of the modern intelligent traffic management system, which has been widely used in many fields. On the Hong Kong-Zhuhai-Macao Bridge, the vehicles may have multiple license plates (LPs) with three different styles, and the traditional contour-based vehicle license plate recognition methods cause a considerable miss rate for multi-style license plates. With such a background, this paper proposes a multi-style license plate recognition method based on feature pyramid network with instance segmentation, which translates the license plate recognition into object instance detection and gets rid of the steps of segmentation and optical character recognition of traditional methods. In the scheme, we design a novel license plate recognition network to precisely locate and classify characters and LP regions concurrently, wherein an assembly layer is added for combining the characters into license plates and outputting license plate strings. The experimental results show that the proposed method achieves 98.57% recognition rate of multi-style LPs on the real world applications. Moreover, we also select the standard license plate datasets, that only contain single style license plates, to test the proposed license plate recognition method, and the corresponding results show that the proposed method achieves competitive performance.