Speed-up Of Connected Component Labeling Algorithm And Its Application In Paper Defect Detection

With the increasing speed of papermaking machines,the requirement for efficient paper defect detection methods becomes higher and higher.Many researchers have paid attention to this subject for a long time.With rapid development of the technologies on image processing and pattern recognition,new ideas and methods for paper defect detection are possibly proposed.For realizing paper defect detection by image processing,it is necessary to recognize independent objects and extract their shape features in an image.Connected component labeling(CCL)on a binary image is one of the most fundamental operations in image analysis,pattern recognition,computer(robot)vision,and machine intelligence.CCL is indispensable whenever a computer or a system needs to recognize independent objects in binary images.Moreover,CCL is necessary for extracting the basic shape features of each object,such as Euler number,area,perimeter,circularity,contour,centroid,etc.For many real-time image recognition systems,e.g.,high-speed on-line detection on an automatic production line,which might need to process dozens or even hundreds of images per second,robot vision,automatic driving,and automatic optical imaging guidance,the performance is directly influenced by the efficiency for recognizing independent objects and extracting their basic shape features.Many algorithms have been proposed for CCL and basic shape feature computation,however,there is still much room for improvement.On the other hand,CCL and basic shape feature computation have been applied to many practical fields,such as machine vision,fingerprint recognition,text recognition,medical image processing,etc.These applications demonstrate that the theoretical research results of CCL have very important practical application values.This thesis will propose several improved algorithms on CCL and basic shape feature computation,and apply them to paper defect detection.The main works of this thesis are as follows:(1)Based on the relationships of neighbor pixels of the current object pixel(s),two multi-row-scan based CCL algorithms are proposed.In conventional CCL algorithms,where pixels are processed one by one in the raster scan,the number of neighbor pixels to be checked for labeling an object pixel and the times for checking the neighbor pixels repeatedly are large,which reduce the efficiency of connected component labeling.Addressing these problems,the proposed algorithms scan multi-row at a time and process multi pixels simultaneously to reduce the average number of neighbor pixels to be checked for processing an object pixel.On the other kind,by taking advantages of configuration transition technology,the proposed algorithms utilize the information obtained during processing the last pixels to reduce the times for checking the neighbor pixels.Thus,the efficiency for connected component labeling is enhanced.The experimental results on various datasets of noise images,medical images,natural images and textural images demonstrated that the proposed algorithms are much more efficient than all conventional CCL algorithms.Moreover,this thesis,from both theoretical analysis and experimental tests,showed that with the increase of the number of pixels to be processed simultaneously,the length of codes for implementing the corresponding algorithm will increase exponentially,which will reduce the execution efficiency of the algorithm.Therefore,the efficient method by scanning multi-row and processing multi pixel could not be unrestrictedly extended.In fact,when scanning four rows at a time and processing four pixels simultaneously,the efficiency of the CCL algorithm would decrease.(2)Research on CCL algorithms for hexagonal pixel images and threedimensional images.Due to the advantages of hexagonal pixel images,a lot of research results on hexagonal pixel images have been publicly published.However,there were few reports on CCL algorithms for hexagonal pixel images.This thesis carried on the research about CCL algorithms for hexagonal pixel images.Based on the research foundation and achievements of CCL algorithms for rectangle pixel images,by analyzing the relationship between the current pixels and neighbor pixels in the mask,a CCL algorithm for hexagonal pixel images,which processes hexagonal pixels two by two,is proposed.Both the analysis of examples and the experimental results on the test data demonstrated that the proposed algorithm is very efficient.On the other hand,when labeling the current foreground voxel in a threedimensional binary image,because the number of neighbor pixels in the mask is large,the number of equivalent label sets that need to be merged is usually large.This thesis proposed a CCL algorithm for three-dimensional binary images,which can merge any number of equivalent label sets by a single merging operation,thus,improve the efficiency of CCL for three-dimensional binary images.(3)By recording and analyzing the relationship between provisional labels and their representative label during the first scan in CCL,an algorithm for combining connected component labeling basic shape feature computing is proposed.In order to reduce the influence on efficiency of image recognition systems caused by multiple scan,the proposed algorithm,during CCL processing,calculates the Euler number,area,perimeter,circularity etc.of each object.Because in the first scan of CCL,all provisional labels in an equivalent label set have the same representative label,the proposed algorithm innovatively calculates area,perimeter,circularity etc.by use of provisional labels.Thus,the proposed algorithm reduces the number of scans from three to two.The correctness of the proposed algorithm is proved and the experimental results demonstrated that the proposed algorithm is much more efficient than the conventional methods on various kinds of images.(4)Based on the above works,a paper defect detection method based on CCL is proposed.Because the gray values and the basic shape feature values of paper defect areas are the key elements for identifying different paper defects,the proposed algorithm directly processes paper defect areas by use of gray images.Thus,the proposed algorithm can omit binarization processing.Moreover,based on the provisional labels in the first scan,the proposed algorithm calculates basic shape feature values of paper defect areas.Therefore,only one scan is needed.The proposed algorithm is simple and easy to be implemented,and could be expected for practical application in paper industry.Through the above works,several meaningful research results have been achieved.These research results can provide new ideas and methods for further improving the efficiency of CCL processing,and theoretical and technical support for improving the performance of real-time image recognition systems.The proposed paper defect detection method is significant for improving paper quality,accelerating independent development of production equipment,and improving the efficiency of our country papermaking enterprises.