License Plate Recognition Algorithm Under Complex Scenes

With the rapid development of intelligent transportation system, license plate recognition system as an important part of it, is becoming more and more important and has various applications. It has been a hot research in the field of pattern recognition, therefore it has a high theoretial research value. Although there are a number of commercial vehicle license recognition systems, but some traditional license plate recognition algorithms are subject to a number of conditions, such as limited scenes, lighting and license plate formats. Robust license plate recognition has enormous research potential.This article aims to research robust license plate recognition algorithms which could work well on different scenes with different image quality, our work focus on the following aspects.1) License plate location. We proposed a rule-based localization algorithm based on edge enhancement combined with domestic mainstream plate formats. We use edge features detection, connective regions scan and regions analysis to achieve an accurate license plate location algorithm.2) Licesne plate character segmentation. We proposed a hybird character segmentation method, which combines connective regions based license character segmentation and CRFs (Conditional Random Fields) based license character segmentation. Connective regions based method is robust,and also has a good segmentent accuracy with low image quality, but unable to segment these license plates whose plate characters are connected with their plate borders. CRFs based method aims to sovle theseplates. It plays a supporting role for connetive regions based method, and highly imporve the character segmentation rate.3) Licesne plate character recognition. We use the SVM classifier to recognize the character image, with gray intensity feature. Models are trained for Chinese characters, numbers and letters to improve recognition accuracy.4) Build a complete license plate recognition system. The overall accuracy rate of experiments on two real datasets is91.83%, higher than average rate of LPR systems. The performance shows that our LPR system has high robustness and enormous practical value.