A High Performance Image Thinning Method And Its Application

In the field of image processing technology,how to extract image shape information features is an important issue.Among them,the characterization,recognition and application of skeleton have attracted wide attention of scholars at home and abroad.Thinning skeleton is an important simplified representation of geometric features of objects.It can maintain the topological features of the original objects to the greatest extent,such as structural connected areas,branching structures,holes or depressions.In image representation and recognition,the results obtained by thinning algorithm have good topology preservation,rotation invariance and other high-performance characteristics,which are of great significance to improve the recognition accuracy.Therefore,how to obtain a high-performance image skeleton that preserves topology,rotation invariance and one-pixel width is a valuable research topic.In this paper,the principles of the reported thinning algorithms,such as distance transformation algorithm,sequential iteration algorithm,rule parallel algorithm and fast parallel algorithm,are summarized.The reasons why these thinning algorithms cannot guarantee the thinning results have the characteristics of maintaining the original image topology,strictly rotating invariant and single-pixel width are analyzed.Combining their advantages,a new thinning algorithm,Ordered Subfield Algorithms(OSA),is proposed.Through theoretical and experimental verification,it is proved that the thinning results of the new algorithm can simultaneously preserve topology,rotation invariance and single-pixel width.The overall performance of the new algorithm is superior to other reported thinning algorithms.Based on the new algorithm,the metal oxide one-dimensional nanorod material image and text image are applied.The results show that the new algorithm can effectively improve the accuracy of nanorod number statistics and the recognition rate of text image,and achieve better thinning effect,thus confirming the practicability and versatility of the new algorithm.The main innovations of this paper are as follows:(1)A new image thinning method is proposed.The ordered subfield method first calculates the Euclidean distance from all foreground elements to the nearest boundary,divides the foreground elements into different subfield according to the ascending order of the distance value,and then assigns a definite operation order to each subfield through the designed subfield scoring(S_subfield)method.In a specific method provided in this paper,the subfield score is calculated by the sum of neighborhood score(S_neigh),centripetal score(S_centri),and normal score(S_norm).Skeleton extraction is a process from low-value subfield to high-value subfield,so all subfields and pixels have a definite operation sequence.Within each subfield,skeleton refinement is processed in parallel,so whether a pixel should be deleted or retained is independent of the processing order.In the case of rotation,as long as the subfield scores remain unchanged,the operation order will not change and the skeleton extracted will not change.(2)The topology of thinning results is kept unchanged by using the method of string and combination.When two elements are operated in parallel,the topological structure of the original graph,especially the connectivity,is often destroyed.By sorting the distance between the elements and the revolving center,this paper divides a pair of elements that destroy the topological structure into different subfield to solve the problem that the two elements cannot decide who to delete first in parallel operation.Because the elements in different subfield are serial processing relations and the elements in the same subfield are symmetrical parallel processing relations,this method ensures that the skeleton extraction results not only have rotation invariance,but also do not change the topological structure of the original graph.(3)In order to solve the problem of strict one-pixel width,this paper first proposes a skeleton classification method,which distinguishes endpoint elements,node elements and branches.Then,on the premise of ordered subfield method,a condition for deleting foreground elements is proposed,which matches the skeleton classification method perfectly.Finally,the branch must have the feature of strict one-pixel width,and the endpoint and node will not.The skeleton that was mistakenly deleted.(4)Several high-standard images thinning performance evaluation methods with strictness and generality are proposed.The new performance evaluation index can accurately verify the advantages and disadvantages of some reported algorithms,and it can also prove that the new method has good image thinning application performance.(5)The new algorithm is applied to the practical application of nanorod quantity statistics.The nanorod image has high density,high overlap and cross-over rate,which makes it difficult to count the number of nanorods.Traditional image thinning algorithms often have the defect of two nodes at X-type intersection in such applications.In this paper,we propose a method to solve the X-type crossover problem,so that the number of end points,nodes and branches can be used to estimate the number of nanorods.The experimental results show that,compared with other thinning algorithms,the end points,nodes and branches obtained by OSA thinning algorithm are closest to the real values,which makes the number of nanorods obtained by statistics have the lowest error rate.Therefore,using OSA can quickly and conveniently estimate the number of rod or line targets with high density and high overlap rate.(6)The new algorithm is applied to the recognition of chess characters.Chess pieces have the characteristics of random placement,uncertainty in the direction of text,and inclination distortion.As a result,the recognition rate of traditional character recognition methods is low.The character recognition methods based on machine learning need large training samples and long training time.To solve these problems,a new method of character recognition based on image thinning algorithm is proposed in this paper.The thinned skeleton is used as a matching template to realize character recognition according to the best matching principle.The new method can recognize text with arbitrary rotation angle and photographs with slight tilt distortion effectively.The experimental results show that OSA has the highest recognition accuracy and the smallest matching template compared with other thinning algorithms because of its strict rotation invariance and single-pixel width.Therefore,OSA can quickly and effectively recognize text characters with rotation and tilt distortion.