Research On MRFS-based Handwritten Chinese Character Recognition

Handwritten Chinese character recognition is an important part of intelligent interfacefor human-computer interaction, which is quite significant for office automation andimprovement of computer’s input efficiency. Due to Chinese characters owing thecharacteristic of numerous categories, complicated structure, similar word, various styles,handwritten Chinese character recognition is one of the difficulties existing in the patternrecognition field. Markov Random Fields encode contextual constraints into the probabilitytheory. Enlightened on MRFs’ successful application in computer vision, this thesis mainlystudied the handwritten Chinese character recognition base on MRFs.Through studying the application of MRFs in solving the problem with complexity anduncertainty, we proposed a MRFs-based statistical structure model and research the MRFs’application for handwritten Chinese character recognition. Compared with the traditionalhandwritten Chinese character model, this proposed model had advantages including:(1)Characteristic of multi-strokes, with which the same stroke of Chinese character could formdifferent MRFs sites according to stroke extraction results. This characteristic skillfullyhandled the uncertainty caused by handwritten style and structure complexity during thestroke extraction process.(2) Model topology formed automatically. Differing fromtraditional method which used a supervised initialization means, the proposed model formedChinese character topology according to neighborhood system in the training stage. Thisinitialization method was more simple.In the process of minimizing MRFs posterior energy fuction in the model matching stage,we presented fuzzy labeling approach for feature sites and a novel optimization strategy byordering of energy fuction to improve the efficiency of the model matching. The experimenton first50Chinese characters in the front of HCL2000database indicated that the proposedmethod for handwritten Chinese character recognition exhibited a recognition accuracy of90.88%, even up to96%for some complex Chinese characters, which fully demonstrated theeffectiveness of the proposed method.In order to keep the stroke connection of character image in the process of strokeextraction, an improved thinning algorithm was put forwarded. The algorithm redefined theconnection constraints in original algorithm and enhanced its adaptability in handwrittencharacter images. The experimental results showed that the character thinning effects has beenimproved by the proposed method and the accuracy of character recognition has increased by8.68%while applied in handwritten Chinese character recognition. The traditional stroke extraction method with Gabor filtering is facing many problemssuch as complex parameter settings and huge template’s computation and so on. To addressthese problems, we came up with a strategy which mixed stroke extraction of Gabor filter andthinning algorithm. The strategy used fast thinning algorithm and reduced the processingworkload of Gabor filter. The experimental results in this thesis presented that the processingworkload of Gabor filter has dropped to6.72%of its original workload.Moreover, the popular handwritten Chinese character feature extraction methods are notconducive to find the subtle discriminative information among similar handwritten Chinesecharacters in subsequent LDA transformation. We suggested a feature optimization methodbased on2DLDA for similar handwritten Chinese character recognition. By representing thetraditional feature extraction and LDA transformation as an optimization task ontwo-dimensional feature matrix of the pixel level features, then optimize the handwrittenChinese character feature matrix by2DLDA. The experimental results on similar handwrittenChinese character recognition indicated that the recognition error rate reduction reached by48.86%by using the optimized method.