Applied Research Of Content Based Image Retrieval(CBIR)in Iicense Plate Recognition System

As an important support system of the intelligent transportation system, the license plate recognition system is of great value in term of application. Different from traditional principle, this paper incorporate the technique of Content Based Image Retrieval (CBIR) into license plate recognition system, to address the problem of how to quickly and accurately recognize single and multiple license plates simultaneously under complex environment.Main contents of the paper include:The first part is to improve the accuracy of the license plate recognition. After an in-depth research on the relevant technology about CBIR, this paper focused on the crucial part---feature extraction in CBIR, which also is of great influence on the efficiency and accuracy of the license plate recognition system. Amongst the image features, local invariant features generally possess high robustness to complex environment, such as SIFT This paper designed a kind of license plate locating method based on SIFT algorithm to adapt the complex environment, and extended the formal character feature database from the pure Chinese characters template to the combination of Chinese characters and Arabic numerals template to eliminate some fake licenses with Chinese characters and improve the accuracy of the locating. Simulational test on the collected106license plate pictures of poor quality show that the accuracy of license locating reached96.23%.The second part is to realize multiple licenses recognition. Based on the feature extraction of SIFT operator, the paper utilized K-Means clustering for coarse locating for multiple vehicle license, and refined the results.Simulational experiments, on locating672valid licenses from the collected221multiple license plate images also show the high accuracy of97.92%.The third part is to speed up the recognition algorithm.This paper firstly replaced the traditional Euclidean distance with a linear combinational distance between City Block distance and Manhattan distance to reduce the calculation. In the matching strategy the paper adopted the improved BBF algorithm based on K-D tree to speed up the matching processionally,the paper evaluated multiple local invariant operators in terms of the robustness and efficiency by the simulational experiments. The results show that SIFT operator and Harris-SIFT operator are superior to other operators.In summary, the contributions of this paper are as follows:(1)Applying the CBIR technology into license plate recognition. Especially, with the local invariant operators, the robustness to complex environment is enhanced.(2)Realization of multiple license plates recognition simultaneously. After clustering the matching features with K-Means algorithm, the center points of the dense region are found for coarse locating the candidate region of the license plate.(3)Evaluation the performance of various local invariant operators in terms of robustness and efficiency.