The Research Of License Plate Recognition System

Nowadays, new traffic rules are becoming more and more strict, which not only warnpeople that they should obey the traffic rules strictly, but also make the requirement of thetopic-license plate recognition effect higher. License plate characteristics in our country arevarious. The article mainly studies the civil cars which are blue plates with white words.License plate recognition system includes: pretreatment, license plate location, tiltcorrection, character segmentation and character recognition. The paper focuses on thecharacter recognition. The implement of the four parts in the front is preparing for the finalrecognition stage, which is also significant in the license plate recognition system. In thepretreatment stage, the plate acquired will be in the process: gray, image enhancement,two-value etc; In the license plate location stage, mathematical morphology and horizontaland vertical projection method and prior knowledge are used in the process; In the tiltcorrection, we use the method based on Radon transform in the process. Considering theneeds of the following work, we conduct the positioning processing based on horizontal andvertical projection after tilt correction. The effect is better. In the character segmentation stage,we conduct horizontal projection and prior knowledge in the process. After the four steps, thecharacters from the experimental results are given the last stage, that is, character recognition.This paper chooses the SVM classifier to classify and it’s divided into three classifiers:Chinese character classifier, the alphabetic character classifier and letters and numbers mixedcharacter classifier. For Chinese characters, recognition experiments select a set of930samples,620group for training,310group for testing; For alphabetic characters, we choose720samples,480group for training,240group for test; For mixed characters, we choose1020samples,680group for training,340group for testing. The SVM parametersoptimization of C and are made by adopting the idea of cross-validation; the training andforecasting of the samples are processed by selecting the appropriate kernel function; at thesame time the comparison with the recognition effect of which does not contain parametersoptimization is conducted under the same training and test set. Recognition rate of SVM afterparameter optimization in Chinese character part is98.710%, in the part of the alphabetic partis99.167%, in the mixed character part reach100%. The overall recognition rate is97.888%and the recognition effect is superior to the no parameter optimization. Experimental results show that the SVM has a good recognition rate and classification generalization ability in thecase of less samples.