A method of automatic image segmentation for gas bubbles in water

Journal of Harbin Instiute of Technology (New Series), Vol. 19, No. 4, 2012A method of automatic image segmentation for gas bubbles in waterLIU Bo, LIN Yan, WANG Yun-long刘波，林焰,王云龙( State Key Laboratory of Structural Analysis for Industrial Equipment，School of Naval Architecture and Ocean Engineering,Dalian University of Technology ，Dalian 116024, China)Abstract: This paper presents an algorithm of automatic bubble image segmentation using the improved ant col-ony optimization methodology. The ant colony optimization method is a metaheuristic algorithm, and has beenapplied in many fields. To reveal the versatility and appropriateness of automatic bubble image segmentation,the fuzzy clustering analysis method is employed in ant colony optimization algorithm. Compared with the well-known image feature extraction operators such as SUSAN and Canny, the proposed method can comparativelysuitable to extract the gas bubbles image edge features. The experimental resuls show that the proposed methodis effective and reliable, and can achieve satisfactory image edge extraction effect.Key words: gas bubbles; image segmentation; ant colony algorithm; fuzzy clustering analysis; watershed algo-CLC number: U66Document code: AArticle ID: 1005-9113(2012)04-0031-06Image segmentation, which is an important issueon, these approaches of edge detection are all proved toin image processing， has been one of the most chal-be successfulHowever, based on these operators oflenging tasks due to the complexity and diversily of im-the edge detection, the image noises are usually mis-ages, and the aim of which is to separate the objecttaken as edges. Therefore how to detect image contourfrom the background and extract edge contour informa-without or with fewer noise's impact would be a majortion. The edge contour generally contains large quanti-task. In addition, none of the foregoing edge detectionties of information about the object, so the edge detec-approaches are generally applicable to all images, andtion often plays an important role in the objects ( under-the different detection methods are usually nol suitablewater targets such as oil platform, ships on the oceanfor application in the same image. In other words ，bottom, etc. ) orientated monitoring, robotic vision andthese approaches, all the time shows their advantagesso on. For the key technology and the present applica-and disadvantages when dealing with different problemstion fields of image segmentation, great interest hasrespectively.been shown in this area-2J and numerous approachesIn view of the above problems, at present a num-have been proposed(3- 61.ber of approaches and their corresponding improve-In general, the algorithm of edge detection!"] is a-ments have been proposed to ensure the accuracy of im-dopted in the experiment of image processing, which isage edge extraction, and some new methods and theo-mainly based on three considerations: firstly, the imageries have been proposed by many researchers, fromcharacters must have invariance in translation, rotationthese edge detection algorithms the readers have evalu-and proportional trans-formation; secondly, these se-ated to determine their relative merits. In recent years,lected features are imaged in the same natural scene ,however, with the further advance of some intelligentand have invariance in different imaging perspectives ;approaches，the methods especially the artificial intelli-thirdly, the Mach effect points out that the human visu-gence ( AI), such as the artificial neural networks, theal is insensitive to the parts of the object in the graygenetic algorithm， the ant colony algorithm and soscale change, and edge is just the part that correspond-orth, have already mushroomed in image processing.ing to the rapidly changing region in image intensity.In this paper, only the ant colony optimization ( ACO)As far as the edge detection is concerned, the tradition-algorithm is introduced for the application of image con-al techniques such as the Log edge detection，the Ro-tour extraction. Due to the ACO method with the dis-berts edge detection, the Prewitte edge detection, thecreteness, good discretion ，robustness , parallel charac-Sobel edge detection，the Canny edge detection, and so ter and po中国煤化I )lars have obtainedReceived 2011-11-15.fYHCNMHGSponsored by the "LiaoningBaiQianW an" Talents Program ( Grant No. 2007-186 -25), the Program of Scientife Research Project of Liaoning Province Ed-ucalion Commission ( Crant No. LS2010046) , and the National Commonweal Industry Scientifie Research Project ( Grant No. 201003024 ).Coresponding author: LIU Bo. E mail: liubochndl@ gmail. com..31.Journal of Harbin Institute of Technology (New Series)，Vol. 19, No. 4, 2012abundant research achievements in applying the ACOputational method to detect bubble shapes in imagesalgorithm to solve the problem of image processing,has a number of interesting applications in the disci-such as image thresholding, image classification， im-pline of oceanography ( analyzing bubble images fromage segmentation and edge detection. An image thresh-the interior of breaking waves， for example， and theolding method using an ACO algorithm has been pro-automated analysis of bubbles from methane seeps) ，posedHl9- 10]，and compared with traditional thresholdingmarine engineering ( the impact of aerated water ontosegmentation approaches, their proposed method canoffshore structures has important design implications )efficiently reduce the calculation time and its results al-and chemical engineering ( bubbles in two-phase flowsso achieved satisfactory thresholding segmentationare photographed and analyzed ). Although the imageseffect. Because of clustering being the unsupervisedof the air bubbles should have been done with actualclassification of patterns ，some researchers combineimages of bubbles in waves or from methane seeps, inclustering with the ACO algorithm for investigating im-order to highlight the superiority of the proposed meth-age classification, this method is more effective and ac-od, the image is acquired under laboratory conditions.curate than the traditional ant colony algorithm'Moreover, for estimating the bubble size, the imageAs far as image edge detection is concerned, Robert etcutting problem of overlapping. bubbles are addressedal. [13] attempted to extract features of leaf outlines andusing the watershed algorithm0l ，and then the nextprimary venation patterns using the basic ant colony al-step is to compute the size of the minor and major axisgorithm in digital images. Although the results achievesemi-diameters of every isolated gas bubble, the aver-satisfactory purposes, some deficiencies still exist, andage diameter, and the coordinate value of the center ofthere are rooms for further improvement. Recently, themass of every bubble.ant colony clustering algorithm gradually becomes a fo-cus for researchers, a fuzzy theory and ACO was pres-2 Overview of Ant Colony Optimizationented to investigate the image edge detection prob-lem', by calculating pheromone fuzzy probability，The ACO algorithm is a kind of bionic evolution,so as to avoid the movement of ants due to the variationwhich was invented by Dorigo and Birattari et al. It isof intensity caused by the noise.inspired by the observation of real ant colonies andThis paper is organized as follows: In Section2, amimiced their finding the shortest route between a foodbrief introduction related research works. Section 3 de-source and their nest. These ants communicate withscribes the ACO algorithm and provides some funda-ach other by chemical agents and receptors, suchmental concepts. Then, an ACO- based image edge de-chemical agents are called the pheromone. Pheromonestection and the analysis of statistical characteristics ofare molecules of glands secreted by the ant' s body andbubbles are proposed in Section 4. After that, Sectiondeposited on the ground when ants walk to and from a5 gives the experimental results of the improved ACOfood source. Ants use pheromone to communicate. Onemethod, which are compared with those of Canny, SU-ant releases pheromone, and other ant can smell phero-SAN algorithm, and presents estimate size of gas bub-mone and follow pheromone with some probability. Ifble. Finally, the imporant conclusion is drawn in Sec-one ant traces a pheromone trail to a food source, thattion 6.trail will be used by many other ants that will reinforcethat trail even more, the more the number of ants trac-Related Worksing this trail, the more attractive this trail it becomes.This autocatalytic process will continue until a trailThis paper investigates anr evolutionary method offrom the ant colony to the food source is established.the ACO algorithm and its application in image edgeThe basic mathematical model of ACO algorithm isextraction. As is known, the ACO is an intelligent ap-summarized as follows:proach'and this metaheuristic is originally de-This method employs artificial ants as simple com-signed to solve the travelling salesman problem, but itputational agents. Set initialize the positions of totallyhas been successfully used in a wide range of applica-ants are m, antk is in city i, and ants select the follow-tions. For example, it was applied to solve the cluste-ing city to be visited through a stochastic mechanism,ring problem191. Different from traditional clusteringthe probability of going lo cityj is given by:methods，the ant colony clustering method has manyI_ _ [r,(l)]°[n,(1)]if j∈allovedkadvantages such as self-governing， flexibility, concur-p:(1)=) S[r(t)1"[n(1)月rency and so forth. Considering these advantages, au中国煤化工tomatic image segmentation of gas bubbles approach inMYHCNMHGotherwise(1)this paper employs clustering theory and ACO algorithmto tackle the image edge detection problem.where allowedk is the set of feasible components，whichThe application of a relatively sophisticated com-is so far constructed the partial solution; The parame-●32●Journal of Harbin Institute of Technology (New Series)，Vol. 19, No. 4, 2012ters a and β control the relative importance of the pher-omone Ty and the heuristic information η;, which is giv-d,=、点pPe(xa-x)2(6)en by:where m is the number of ants; p is the weight coeffi-η。=j(2)cient.Secondly, computing pheromone quantity phy, setwhere dy is the distance between cities i andj.r is the clustering radius, and the formula is definedAt each iteration, the pheromone values are upda-below :ted by all the m ants have been building a solution inthe iteration itself. The pheromone T，associates withphy={lifd;≤r(7)otherwisethe trail joining cities i andj, is updated as follows:Thirdly, we compute the transition probability ofrg(l+n) = (1 -ρ)r;(l) +△r;(t) (3)merging every pixel X; to thresholdr for the kth ant, itsAr;() =二" Or(t)(4)formula definition as follows:whereρ is the evaporation rate; Org(t) is the quantity[phg(t)]g[n,(1)]Pifj∈allowed;of pheromone laid on the trail (i;,j) by ant h.p()={_ 2 [l()]"n.()]PAccording to the different pheromone update strat-salouedkegy，Dorigo proposed three different models of basicant colony algorithms, called the Ant-Cycle model, theAnt-Quantity model and the Ant-Density model, re-where a and β are the parameters that reflect the apoca-spectively. The difference is that the△rj(l) of differ-lyptic factor and the expected apocalyptic factor, re-ent calculation methods. This text describes the Ant-spectively; the quantity of visibility guidance functionCycle model, and the△rj(t) is calculated as follows:η; equals to一, which denotes the heuristic value whenAr(1) =if k-th ant use tail(ij) (5)ant moving from pixelX; and pixelX, allowed; = {X,“I d,≤r,s =1,2,.,N} is the set of all availablewhere Q is a parameter that speifies the amount ofpaths.pheromones distributed by an ant h, and Lp is the tourFourthly, with ant moving and completing a cy-length of the ant k.cle, the pheromone amount on every path can be ad-The ACO algorithm becomes a promising researchjusted; the pheromone updated is adopted according toarea and further improved algorithms as Ant Colonythe following formula:System ( ACS)，Max-Min Ant System and so forth. Byph;(i+1) = (1 -p)phg(t) + Oph;(I) (9)accurate definition of the problem, the ACO methodAphg(t) = 2n Aphg(t) .(10)can be applied in various fields.whereρ is a pheromone evaporation ceofficient; (1 -ρ) represents the residue factor of trail between th3 Proposed Image Feature Extraction Approachtimetandt+ 1. Ophy(l) is the pheromone quantity in-First of all, this paper presents an improved meth-crement by ants in the edge trail(i,j); Sph;(t) is theod of edge detection based on the basic ACO algorithm.pheromone quantity increment left by the hth ant in theAfter that, the estimated dimension of the bubble isedge trail(i,j).Fifthly, due to the travel of ants are random andcomputed on the basis of this research in this section.This method takes X for an original image, and X;(i =blind, such that the time of image feature extraction1 ,2,.",n) stands for each pixel of the image, and eachwill be sluggish. In this paper, the fuzzy clusteringpixel is as an ant that represents a character vector ofmeans ( FCM) algorithm is adopted to solve thisgray-value，gradient and neighborhood. The contourproblem. For theX;(i = 1,2,",n), they are classi-points on the image are the foraging food resource byfied asc fuzzy field of subset, that Si,S,",S., ifuyants, When going from pixel i to pixelj, each ant X{(idenotes the degree of x; belongs to S, and then the= 1,2,..,n) lays pheromone ph, on trail(i,j). Eachfuzzy C-set is theU = {u;| 1≤i≤c,1≤j≤n}, inant according to a transition probability P; in path selec-which2uy=1,1