Research On Thinning Algorithm Of Chinese Characters

With the rapid development of computer technology,Chinese character recognition technology is well studied,and it has been applied to many fields,such as intelligent sorting of mails,manuscript checking,handwriting identification,check and invoice processing and so on.As well,the thinning algorithm is one of the most important techniques in Chinese character recognition,and it is also the bottleneck of the off-line Chinese character recognition.In this paper,we focus on learning and doing experiments to compare the existing thinning algorithms,and analyze the advantages and disadvantages of the existing algorithms and then make some improvements,the main works of the paper are as following:(1)The preprocessing of image before thinning are studied.We focus on the minimum bounding box algorithm based on the two methods of vertical and horizontal integration method and region expansion method.In the vertical and horizontal integration method,the wavelet transform is used to smooth the integral curve,which improves the accuracy of column and character segmentation.Experiments show that the accurate implementation of the character segmentation is obtained by using the methods.Several thinning algorithms were studied,including Z-S thinning algorithm,the Maximum Inscribed Circle skeleton extraction algorithm,thinning algorithm based on mathematical morphology,thinning algorithm based on index table and A-W thinning algorithm etc,which focus on realizing these algorithms and comparing their advantages and disadvantages.(2)Although the A-W thinning algorithm is better than the other methods,there are still some problems,such as the non-single pixel,fork point distortion,spurious branches etc.Single pixel processing algorithm based on adjacency matrix is not good at processing the stroke slash and back slash.Single pixel processing algorithm based on template matching is good at processing the stroke slash and back slash except the crossing.We put forward the new method CEADAR(Combination of template matching and adjacency matrix)to eliminate the two pixel wide skeleton and achieve good results.Then,due to the shortage of A-W thinning algorithm,an improved thinning algorithm is proposed on the basis of A-W thinning algorithm.This algorithm combines the A-W thinning algorithm with the stroke trend prediction and preferentially retains the horizontal and vertical,slash,back slash,the four strokes of the points.The experimental results show that this algorithm improves the distortion phenomenon at the cross of strokes and fine results can be obtained.(3)A distortion correction algorithm is proposed.The feature points are detected on the skeleton.In view of the phenomenon that the skeleton by using the maximum circle algorithm is not smooth after stroke connection.Then we employ the improved maximum circle method to merge the fork points and reconnect the local strokes.The connection of the local stroke is optimized with the knowledge rule of Chinese character structure,which can obtain better results.The experimental results show that the refinement results obtained by this algorithm are single pixel,without burr or bifurcation point,and good results are obtained on the dataset of calligraphy and Song,Kai,Hei,Li styles.