A Vehicle License Plate Segmentation Based on Likeliest Character Region

This paper presents a novel scheme towards license plate segmentation. Due to various illumination conditions, an adaptive local binary method is performed to acquire binary images. Then likeliest character region (LCR) is detected based on canny edge images which are acquired from preliminarily processed license plates. All binary images are transformed into a standard positive image without recognizing colors of plates. Next, the energy evaluation function is put forward based on conventional projection histograms of positive images. Eventually, remaining characters can easily be searched using LCR as well as the energy evaluation function even if the characters are partially touching to each other or heavily contaminated. From the experiment, we obtained encouraging result of 97.2% on our own challenging database, taken from real scene under different conditions in China. Results also demonstrate that our method greatly outperforms the conventional methods.