Analysis And Research Of Handwritten Character Clustering Based On Affinity Propagation Clustering Algorithm

Character recognition has always been a pop research issue in the pattern recognition field. It has important application value and social implication in the digital office automation, word information storage and other fields. At present, the most digital character recognition technology has been applied in our daily life. However, the research of off-line handwritten character recognition has still been the stage of experimental study currently. Handwritten characters have had many characteristics like high similarity, unconstrained writing, and multiple writing styles for one character, which have seriously impacted the recognition quality of off-line handwritten character.Handwritten Chinese character was the research object of the thesis. And then, off-line handwritten Chinese character recognition system platform was established on the basis of traditional pattern recognition model. Every step in the platform from character image standardization to statistical feature extraction and then high dimension feature compression to classifier training and generating the character dictionary for the recognition of test samples. Aiming at the problem that using single character template obtained low recognition rate in the system, the advantage of the clustering analysis that it made reasonable classification according to the different forms for the same object was applied in the recognition system. The low dimensional statistic feature of all the training sample for the same Chinese character was made clustering to obtain recognition dictionary with multiple templates. The experiment displays that clustering algorithms can improve the recognition accuracy of the off-line handwritten Chinese character recognition system. And Affinity Propagation(AP) clustering algorithm can achieve preferable result.For the defect that AP clustering needs to set up preference function to obtain different classified number, a modified AP clustering algorithm was proposed. The characteristic that clustering judging function can evaluate clustering quality was applied in the modified algorithm. Class number was updated by changing preference parameter adaptively, and then the optimal clustering result was obtained by assessing clustering quality of every iteration. The experiment indicates that the modified AP clustering improves the convergence and clustering quality, and has the best performance in the off-line handwritten Chinese character recognition.