ABSTRACT
License plate extraction is considered to be the most crucial step of Automatic license plate recognition (ALPR) system. In this paper, a region-based license plate hybrid detection method is proposed to solve practical problems under complex background in which existing large quantity of disturbing information. In this method, coarse license plate location is carried out firstly to get the head part of a vehicle. Then a new Fast Mean Shift method based on random sampling of Kernel Density Estimate (KDE) is adopted to segment the color vehicle images, in order to get candidate license plate regions. The remarkable speed-up it brings makes Mean Shift segmentation more suitable for this application. Feature extraction and classification is used to accurately separate license plate from other candidate regions. At last, tilted license plate regulation is used for future recognition steps.
1. INTRODUCTION
In recent years, Intelligent Transportation Systems (ITS) has been a priority and hotspot issue in the research field . Use the image processing technology to analyze the vehicles is an important part in the ITS. The most effective way to identify the vehicles is to tell the license plates apart. It mainly includes three techniques: Plate Location, Character Extraction, Character Recognition. Vehicle license plate detection is one of the most important part in vehicle recognition. The accuracy and speed of the location will directly affect the performance of the overall system.
6. CONCLUSION
Complex backgrounds and various scenes bring lots of difficulties to ALPR researches. Considering this situation, this paper presents a region-based method based on the new Fast Mean Shift algorithm. Compare with standard Mean Shift’s calculation complexity, this accelerated method make color image segmentation more suitable for this application which has a strict speed requirement. In the experiment image library, there are about 400 images, size about 640×480, including various conditions, such as strong lighting, degrading of license plate, different view of shooting. The result demonstrates our method works well under these interferences, and the accuracy of detection is 92.6%. This is because the candidate regions are divided directly from color images rather than generated from features as most edge-based methods do. Our future works will focus on improving precision of the location algorithm and further optimization of the program for better efficiency.
