License plate recognition based on SIFT feature

Abstract

Although license plate recognition (LPR) system is widely applied in practice, its some key techniques that cannot meet application requirements in natural scenes still need more attention, such as Chinese character recognition, candidate filtration, tilt correction, and character segmentation. In this paper, a novel method based on SIFT feature is devised to solve the four problems simultaneously. Promising experimental results demonstrate that the proposed is robust to various adverse factors, such as complex background, scaling variation, rather large tilting angle, contamination, illumination variation, partial occlusion, and defective character. The success rates of Chinese character recognition and candidate filtration reaches to 96%; the tilt correction accuracies reach to 0.177° and 0.238° in horizontal and vertical directions respectively; and the success rate of character segmentation approaches 100%. Remarkably, the average execution time for these four processes is lower than 268ms, which may favor real-time processing.