Research On The Application Of Affinity Propagation Clustering For License Plate Character Segmentation

Character segmentation is a key problem in the fields of automatic license plate recognition.Even though this topic has been widely researched,there still exist many problems need to further explore.Nowadays,the main methods of license plate character segmentation algorithm are region partitioning,K-means clustering algorithm,fuzzy C means clustering algorithm.However,those methods have some problem such as incomplete segmentation,excessive segmentation problems and so on,which would decrease the accuracy of character segmentation.Therefore,it is significant to explore novel approaches to improve the accuracy and effectiveness of license plate character segmentation.In recent years,using affinity propagation(AP)clustering algorithm in license plate character segmentation algorithm has become a hot area of research.Compared with those license plate character segmentation methods based on region partitioning,K-means clustering algorithm,fuzzy C means clustering algorithm,it has better adaptability that it does not set the number of clustering and its centers in advance.This study has come up with a new method of license plate character segmentation based on the modified AP clustering algorithm.The merits of this method are lists as below:(1)The AP clustering algorithm is a kind of unsupervised clustering(UC).It dose not need to initializes the number of cluster centers,but treat each data point as a cluster center,and the optimal number of cluster centers is obtained by iterating.Therefore,it can solve the issues that the existing methods of license plate character segmentation such as those methods based on K means clustering algorithm,fuzzy C means(FCM)clustering algorithm,must preset the amount of the character in a license plate image.(2)The quantity of recognized exemplars(quantity of clusters)is greatly influenced by the values of the input preferences(p value)in the AP clustering algorithm,and the p value of the AP algorithm is equal to themean of the values on the diagonal of the similarity matrix.It does not take into account the effect of extreme values which may led to the selection of clustering center is not reasonable.The improved AP clustering algorithm has considered the effect of extreme value on the number of cluster centers by selecting p value according to the median number so that improve accuracy of the selection of clustering center.This study captured license plate image by the Matlab Experiment Platform.In order to compare the accuracy rate of license plate character segmentation methods based on K-means clustering algorithm,fuzzy C means clustering algorithm,classic AP clustering algorithm and the modified AP clustering algorithm,All of those images were segmented by each method.Results showed accuracy rates of those four methods for ordinary license plate character segmentation were 79.2%,88.2%,90.09%,94.3%,and for multilayer one were 55.5%,66.6%,69.4%,88.8%,respectively.By use the four clustering algorithms to clustering the iris and other international standard datasets,we obtained the mean and variance of the Partition Coefficient and Classification Entropy.It indicated that the application of the modified AP clustering algorithm for license plate character segmentation can improved the accuracy and effectiveness of license plate character segmentation.