An Efficient and Unified Recognition Method for Multiple License Plates in Unconstrained Scenarios

Abstract

License plate recognition is an important research topic in image recognition, which plays a vital role in helping the government effectively manage vehicles. At present, most license plate recognition methods focus on a single category and specific scenes. The requirements on distance, illumination, angle and other conditions are pretty high, which are not conducive to the application in practice. This paper proposes a real-time and accurate automatic license plate recognition (ALPR) method to overcome these challenges. The ALPR method contains a license plate detection network (LPDNet) and character recognition network (CRNet). LPDNet is based on the anchor-free method to detect the bounding box and four corners of license plates in an unconstrained environment. The proposed Rotating Gaussian Kernel and centrality loss enable LPDNet to improve performance in complex environments. CRNet is composed of a full convolution network, which can identify many kinds of license plates. In this network, feature extraction and compression technology have been proposed, effectively identifying multiple license plates in combination with a Character Classification module. Experiments on multiple public datasets show that the method outperforms existing methods in speed and accuracy, especially in complex and challenging environments.