| With the rapid development of artificial intelligence technology and the rise of “Made in China”,the recognition of Chinese characters has attracted more attention.Among them,the feature information such as the end points,intersections,and angles between pixels obtained by handwritten Chinese character refinement is a part of the basis for handwritten Chinese character recognition.This paper focuses on the basic principles of the existing AW,Euclidean distance,K3 M,and neural network thinning algorithms,and compares the advantages and disadvantages.Combining the rotation invariance of AW algorithm and the easy realization of K3 M algorithm,a new thinning algorithm is proposed.The BP network model based on new thinning rules was built on the LabVIEW platform to realize the thinning of handwritten Chinese characters,and the thinning effect of neural network algorithms under different neighborhood sizes were studied.The main work of the paper is as follows.(1)Analysis of the image preprocessing process on the LabVIEW platform found that the preprocessed image border pixels will produce mutations,forming burr points or pits,affecting the thinning effect.In view of this situation,the image pruning method based on template matching is proposed to repair,reduce the impact of burr points and dents,and also has good noise immunity.(2)Improvement of the K3 M algorithm: a)using the distance field to mark the boundary of the image,which can effectively solve the problem of duplicate mark existing in the original K3 M algorithm,effectively reduce the time consumption,and make the refined result have better middle axis;b)analyze the template of the query array and find that all of them are based on 4-connected templates.In addition,there are redundant branches and non-single-pixel width problems.The query array is reconstructed by improved template,which reduces the redundant branchs,and has better 8-connected and maintains the original topology.(3)Propose a new thinning algorithm: a)combine the rotation invariant template of AW algorithm,optimize the combination of the improved templates,and construct an optimized combination template that maintains the original advantages and has rotation invariance;b)combine the stroke direction of handwritten Chinese characters to improve their optimized combination template,and solve the problem of double pixels such as two-pixel diagonal lines and 2×2 square.The results show that the thinning rate is better than the K3 M and Euclidean distance thinning algorithms.(4)Combined with the development of artificial neural network algorithm,the BP neural network module with strong learning ability and generalization ability is built on the LabVIEW platform to perform new thinning rule learning,and it forms a thinning system with preprocessing module to realize skeleton extraction of handwritten Chinese characters.At the same time,it is studied that under the same new thinning rules,the thinning effect of BP algorithm with different template sizes is basically the same,and it can be determined that the thinning rule is more important than the template size.The experimental results show that the system is stable,and can effectively obtain a skeleton with good connectivity,rotation invariance,and neutrality.It also has the characteristics of single pixel width and less burr. |
