Elsevier Editorial System(tm) for International Journal of Machine Tools and Manufacture Manuscript Draft Manuscript Number: Title: On-line tool wear monitoring using geometric descriptors from digital images Article Type: Research Paper Section/Category: Keywords: Tool wear, monitoring, computer vision, image classification Corresponding Author: Dr. J. Barreiro Garcia, Corresponding Author's Institution: University of Leon First Author: M. Castejón, Doctor Order of Authors: M. Castejón, Doctor; E. Alegre, Doctor; J. Barreiro Garcia; L. K. Hernández, Master Manuscript Region of Origin: Abstract: A new method based on a computer vision and statistical learning system is proposed to estimate the wear level in cutting inserts and to identify the time for its replacement. AISI SAE 1045 and 4140 steel bars of 250 mm of length and 90 mm of diameter were machined using a CNC parallel lathe. The image acquisition system comprised a Pulnix PE2015 B/W camera; a 70XL industrial zoom, with an extension tube of 1X; several lenses, a DCR®III regulated light source and a diffuse lighting system. The images were captured by a Matrox Meteor II card and pre-processed and segmented with Matlab. For each wear region, a set of 9 geometrical descriptors was obtained. The cluster analysis revealed the presence of three distinct categories that corresponded to low, medium and high wear levels. The effectiveness of the classification was verified by means of a LDA class reconstruction that reported a Fowlkes-Mallows index of 0.8571. The LDA likelihood estimates of the wear region provide a useful tool insert replacement criterion. * Title page with author details Title: On-line tool wear monitoring using geometric descriptors from digital images Abbreviated title: Tool wear monitoring using geometric descriptors M. Castejóna, E. Alegrea, J. Barreiroa, 1, L.K. Hernándezb a Universidad de León. Escuela de Ingenierías Industrial e Informática, Campus de Vegazana. 24071. León. SPAIN b Universidad de Pamplona. Dpt. de Ingeniería Mecánica, Industrial y Mecatrónica, Km. 1 vía Bucaramanga. Pamplona. COLOMBIA. 1 * Corresponding author. E-mail address: dfqjbg@unileon.es Tel. +34 987 191792 Fax: +34 987 291930 Manuscript without author identifiers ABSTRACT A new method based on a computer vision and statistical learning system is proposed to estimate the wear level in cutting inserts and to identify the time for its replacement. AISI SAE 1045 and 4140 steel bars of 250 mm of length and 90 mm of diameter were machined using a CNC parallel lathe. The image acquisition system comprised a Pulnix PE2015 B/W camera; a 70XL industrial zoom, with an extension tube of 1X; several lenses, a DCR®III regulated light source and a diffuse lighting system. The images were captured by a Matrox Meteor II card and pre-processed and segmented with Matlab. For each wear region, a set of 9 geometrical descriptors was obtained. The cluster analysis revealed the presence of three distinct categories that corresponded to low, medium and high wear levels. The effectiveness of the classification was verified by means of a LDA class reconstruction that reported a FowlkesMallows index of 0.8571. The LDA likelihood estimates of the wear region provide a useful tool insert replacement criterion. KEYWORDS Tool wear, monitoring, computer vision, image classification 1. INTRODUCTION The development of on-line measurement systems to report the wear level of tool inserts in automated metal cutting is an issue of paramount importance for the control of automated production environments. The production costs cut down associated to lower human resources requirements ---the supervision of a machinist is no longer required--- and the benefits of operating at higher cutting speeds come, unfortunately, at the price of shorter tool lives. Last decade has witnessed a growing awareness of the increased production benefits related to extending the use of the tool insert beyond the standard limits. In spite of the many efforts focused on this issue, the quest for a satisfactory on-line monitoring solution has not yet reached an end because of the great difficulties involved in wear measurement. Turning, milling and drilling [1] are processes involved in the manufacture of the large majority of goods, which makes the cost of the tool inserts become an important amount of the items production costs, especially in the unmanned production context. As reported by Teti [2] and Weckenmann et al. [3], the costs of the cutting tools and their replacement account for three to twelve percent of the total production costs. Though small at first glance, these permanent costs constantly accumulate along the machine life cycle [4]. In addition, about twenty percent of non-productive time is due, in modern machines, to tool failure [5]. On-line monitoring of the wear level might improve the decision making process involved in tool insert replacement by means of taking advantage of a based on 2 facts approach instead of relying on the subjective criteria of the machinist. Such an approach would lead to a forty percent shrink in production costs [6], a clear business opportunity. It is very important, therefore, to develop precise methods to predict the tool wear level, its evolution and remaining useful life. The machine control system should be able to provide optimized strategies for tool replacement and the adjustment of tool correctors. Tool wear measurement studies carried out so far provide interesting information about the tool condition during the machining operation; however, the results are limited to the measurement of very particular dimensional characteristics [1][3][7-11]. It would be interesting to complete this information with more features to better determine the actual state of the tool. Other researchers have combined signals coming from different types of sensors, which is known as sensor fusion [12-15]. However, the bulk of the literature in this context propose the use of indirect measures obtained from acoustic and vibration sensors, cutting forces, power or current consumption and others. It would be adequate as well to consider direct wear measurement so as to improve the performance. Wear and its measurement are described in several standards (ISO 3685, ISO 8688, ISO 883, ISO 3364, ISO 6987 and ISO 9361). These standards provide wear threshold values at several points of the worn region as a reference to replace the tool insert. However, the technological advances in machines, tools and part materials have rendered these standards obsolete. The threshold 3 values included in these standards are too conservative for current technology and, thus, their industrial application is limited. Many studies demonstrate that machining in good conditions is still possible over these threshold values. Besides this, practical rules for tool replacement habitually used at industrial environments are not good enough, since they are based on the planned values of cutting time, number of parts produced and similar. These rules do not optimize tool life. For these reasons, during last years several researchers have been working in the qualitative and quantitative morphology determination of tool wear, as well as in tool monitoring systems. State of the art in sensor systems and their industrial application has been investigated by diverse researchers [5][16-22]. These works classify measurement techniques in two main groups: direct and indirect methods. Direct methods measure the wear directly on the tool, while indirect methods use intermediate process variables that are correlated with the tool wear, such as cutting forces, acoustic and vibration signals and others. In general, indirect methods ease the tool wear measurement, since they do not involve a process stop. For this reason, they have been the most popular methods for years. However, the resulting precision is not as good as that achieved with direct methods, since the measurement is affected by noise signals [5][9][21][23]. Therefore the desire for the implementation of direct methods for wear measurement without machining stops and no tool extraction from the machine or tool-holder. 4 A review of monitoring systems is showed in [24]. This paper indicates that direct optic systems are the most reliable, in spite of their high cost. Nevertheless, continuous progress has taken place in sensors technology and, in particular, vision sensors have been specially improved in performance but also in cost reduction. For this reason, different researchers have used them for on-line tool wear measurement. In addition, advances in image processing and artificial intelligence technology provide more reliable image analyses and tool replacement strategies. The high number of applications cited in the state of the art demonstrates that computer vision systems are already a mature technology. However, the identification of different wear morphologies has not been efficiently achieved yet. This is one of our goals in this research. State of the art in tool condition monitoring (TCM) using a computer vision system indicates two work lines: i) systems with directional light for flank wear measurement [1][5][8-10][25-27] and ii) systems with structured light and 3D reconstruction for crater wear measurement [3][23][27]. Most of the industrial applications indicate that the predominant defect that determines tool substitution is the excessive wear in flank, in contrast with the wear in crater that is restricted to particular part material - tool material matching. In consequence, our research is focused in flank wear characterization of turning tool inserts by means of computer vision techniques together with classification techniques based on geometric descriptors. The objective is to identify the morphology and wear level to adopt the corrective actions in each state. 2. MATERIALS AND METHODS 2.1. Machining and vision systems A CNC parallel lathe has been used for the machining with a maximum turning speed of 2300 rpm. AISI SAE 1045 (20 HB, normalized) and 4140 (34 HB, tempered) steel bars of 250 mm of length and 90 mm of diameter were machined. The tool inserts were made of covered tungsten carbide, rhombic, high tough. Different values were used for the cutting parameters: cutting speeds (Vc) between 140 and 200 m/min, feed rate (f) of 0,2 mm/rev and depth of cutting (ap) at 2 mm remain constant. The tool is disassembled and the insert is located in a tool-fixture after machining a pass along the part; tool-fixture allows keeping constant the flank location in the image. Additionally, measuring of roughness and hardness was done on the machined surface [30]. 2.2. Image acquisition Images have been acquired [30] using a Pullnix PE2015 B/W camera with 1/3" CCD. Digitalization was carried out with a Matrox Meteor II card. The optical system is composed by a 70XL industrial zoom of OPTEM, with an extension tube of 1X and 0.5X/0,75X/1.5X/2.0X OPTEM lens. The lighting system is composed by a DCR®III regulated light source of FOSTEC that provides an intense cold lighting. A SCDI system of diffuse lighting of NER SCDI-25-F0 is used to avoid shines. The system provides diffuse lighting in the same direction 6 as the camera axis. Positioning of lighting is carried out by means of a single bundle of Fostec. Figure 1 shows the machine with the vision system. Image acquisition is achieved using a developed application that uses the Matlab Image Acquisition Toolbox. The application has three main modules: setup of the camera, setup of the sequence and acquisition of the image. These modules provide information of the capturing device and let the user choose the resolution, define the information storage path and save the images taken. 2.3. Image processing With the described vision system, 1383 flank images were acquired. Initially, a low-pass filter was applied to the image for background blurring and to make easier the segmentation. Later on, the region of interest is cropped and the contrast is enhanced by means of a histogram stretching. After that, the images were segmented manually. The original images, in grey scale, and the binary segmented images were stored in tiff format with compression without loss (figure 2). Image processing and geometrical descriptors extraction were done using Matlab and the Image Processing Toolbox. 2.4. Image descriptors The following geometrical descriptors were obtained for each of the 1383 binary images which resulted from the image segmentation: a) Area: number of pixels in the wear region. b) Length of major axis: length of the major axis of the ellipse that has the same normalized second central moments as the region. c) Length of minor axis: length of the minor axis of the ellipse that has the same normalized second central moments as the region. d) Eccentricity: eccentricity of the ellipse that has the same second moments as the region. The eccentricity is the ratio of distance between the ellipse foci, c, and its major axis length, α (1). Value is in the range 0 to 1 (0 and 1 are degenerated cases; an ellipse with eccentricity 0 is a circle, while an ellipse with eccentricity 1 is a line segment). e= α c (1) e) Orientation: angle in degrees between the x-axis and the major axis of the ellipse that has the same second moments as the region. f) Convex area: number of pixels in the Convex Image. Convex Image is the convex hull with all pixels within the hull filled in. The convex hull of a set of points S in n dimensions is the intersection of all convex sets containing S. For N points P1, …, PN, the convex hull C is given by the expression (2). ⎧N c = ⎨∑ λ j p j :λ j ≥ 0 ∀ j and ⎩ j =1 ∑λ N j =1 j ⎫ =1⎬ ⎭ (2) g) Equivalent diameter: diameter of a circle with the same area as the region. It is computed as indicated in (3). 8 ED = 4 ⋅ Area π (3) h) Solidity: proportion of pixels in the convex hull that are also in the region. It is computed as indicated in (4). S= Area Convex Area (4) i) Extent: proportion of pixels in the bounding box that are also in the region. It is computed as indicated in (5). The Bounding Box is the smallest rectangle containing the region. E= Area Bounding Box Area (5) A vector of features is generated. It is composed by the nine previous values and describes each of the wear regions. 3. TESTS AND RESULTS 3.1. Determination of the existing wear classes The collected features allowed us to perform a cluster analysis in order to reveal the underlying structures supporting the data set. A finite mixture model approach was adopted, in consideration of the complex nature of the data. For that purpose, we found extremely useful the implementation of the Mclust algorithm [31-35] running under R, a free software environment for statistical computing and graphics [36]. The obtained results suggested the presence of three clusters that could, eventually, be referred to three distinct wear levels: low, medium and high. The Mclust algorithm provides the mixture of multivariate normal models that best fit the supporting data. Along the way, Mclust takes advantage of the EM (Expectation Maximization) algorithm [37,38] in order to find the maximum likelihood set of parameters. The results comprehend not only the optimal values of the parameters but also some hints about the number of clusters by means of the Bayesian Information Criterion [39]. Whereas traditional clustering techniques perform a hard partitioning of the data set, the finite mixture model approach allows each observation - each feature vector describing the worn region - to belong simultaneously to every cluster. The degree in which an observation belongs to a certain class depends on its proximity to the center of the class, a distance that is measured according to the defining parameters of each cluster. Of special interest is the linear discriminant analysis (LDA) [40,41] projection of the data set obtained by considering the class identifiers resulting from the Mclust algorithm, which is showed in figure 3. The three classes can be better distinguished by virtue of the boundaries amongst wear clusters that divide the map in three regions: low worn region (left side), medium worn region (center) and high worn region (right side). The LDA approach allows to test the quality of the results provided by the clustering algorithm; on the basis of considering that a clustering has been 10 correctly determined if their parameters can be used back again to build a predictive model whose results mimic the original behaviour. The 1.386 wear observations acquired were classified according to LDA as follows: 33,91% were assigned to class L (low wear); 45,40% to class M (medium wear) ; 20,67% to class H (high wear). The comparison amongst the results obtained by the Mclust algorithm and those obtained by the prediction performed using the discriminant analysis can be expressed as a confusion matrix [42], as shown in table 1. The cij element of the matrix represents the number of elements that simultaneously belong to the Mi and Lj classes, where Mi is the i-th class identified by the mclust algorithm and Lj is the j-th class predicted by the LDA. The information contained in table 1 can be summarized in a single index following the procedures described by Fowlkes Mallows [43,44]. The FowlkesMallows index resulting from table 1 is 0,8571, a high value that supports the results of the Mclust cluster analysis. The LDA projection shows as well the evolution of the wearing process as a continuum through which the insert tool makes its way from the outset in the low wearing degree region, to the medium and high wearing degrees. 3.2. Significant features We also used LDA analysis to select the features from the vector that are more significant for the classification. It was detected that a 98,63% of the information necessary to do the classification could be obtained using three of the nine descriptors. These descriptors are eccentricity, extent and solidity. Table 2 shows the weights of the descriptors in each of the LDA projections made. 3.3. Determination of tool life The discriminant analysis provides not only an insightful projection of the tool insert condition onto the wear map, but also the wear evolution followed by each insert during the machining. This evolution represents the likelihood of belonging to the dominant class at the current wear map location. This estimate evaluates the wear level of the insert and, therefore, might be of help in the development of a tool replacement criterion in order to preserve part tolerances: should the wear level be at the end of the class M (medium wear), the tool must be replaced to avoid the entry in class H (high wear). The evolution of the wearing process can be clearly traced onto the wear map obtained by the LDA projection, as figure 4 shows. This figure locate the wear condition of 14 images taken for an insert. The sequence corresponds, from the left to the right, to an increment in machining time and therefore in tool wear. The first four observations belong to class L (low wear), the following five are in class M (medium wear) and the last five observations are in class H (high wear). Figure 5 shows the probability of a wear state to belong to each of the three classes established. This figure contains the same observations as figure 4, since it corresponds to the same insert. It can be appreciated that the three first 12 observations have a probability of belonging to class L higher than 95%. The forth and fifth observation show a probability of belonging to classes L and M higher than 85%. A similar behaviour can be observed in the transition region between classes M and H; the tenth observation corresponds to class M, but the probability is lower than 85%; the following observation is included in class 3 but the probability is lower than 75%. The last three observations belong to class 3 and were acquired before tool replacement due to not practical machining conditions. This behaviour is present in every insert tested and allows establishing a tool life criterion. If a conservative criterion is used, tool life finishes when the observation reaches values of probability of belonging to class M lower than 85%. Should a less restrictive criterion be desired, the tool insert must be replaced before the probability of belonging to class H surpasses the 80% value. Our tests were done at high cutting speeds to accelerate the wear of inserts. A drawback of this procedure is that the amount of acquired images were not enough to provide a smooth evolution, which sometimes produced an abrupt transition among classes. In standard machining, where a slower rate of wear is expected, it is possible to use the proposed criterion with the certainty that machining with highly worn tools will be prevented. 4. CONCLUSIONS AND FUTURE WORKS We have developed a method to establish the wear level of tool cutting inserts. Our proposal is based on the analysis of the geometrical features of the worn region image. A computer vision system procures these images as the machine operates and the discriminant analysis of the geometrical features estimates the wear level of the tool insert. Should it be superior to a defined threshold, the monitoring system alerts for its replacement. Such an approach leads to a more sensible utilization of tool inserts to the benefit of the manufacturing system as a whole. The performed analyses provided insights into the monitoring of the wear process. The linear discriminant analysis based on the wear clusters distinguished a subset of three out of nine descriptors that accounted for more than ninety eight percent of the information needed to establish the wear condition. These were the eccentricity, extent and solidity shape descriptors. Moreover, the reliability of the model based cluster analysis ---assessed by means of the Fowlkes-Mallows index--- and the correspondence of the resulting clusters to three distinct wear levels underpinned our approach. Future work include automating the images segmentation process for delimiting the wear region as well as considering other shape descriptors. The former will ease the industrial implementation of our approach while the latter will further improve our understanding of the wearing process. 14 ACKNOWLEDGMENTS This work has been possible thanks to the financial support of the Junta de Castilla y León under grant No LE018B06, the University of León under grant No ULE2005-01, and the Spanish Ministerio de Educación y Ciencia under grants No DPI2006-02550 and DPI2006-14784. REFERENCES 1. M. Sortino, Application of statistical filtering for optical detection of tool wear, Int. J. Machine Tools & Manufacture 43 (2003) 493-497 2. R. Teti, Machining of composite materials, Annals of the CIRP 51 (2002) 611634 3. A. Weckenmann, K. Nalbantic, Precision measurement of cutting tools with two matched optical 3D-sensors, Annals of CIRP 52 (2003) 443-446 4. J. Kopac, Influence of cutting material and coating on tool quality and tool life, J. Materials Processing Technology 78 (1998) 95-103 5. S. Kurada, C. Bradley, A review of machine vision sensors for tool condition monitoring, Computers in Industry 34 (1997) 55-72 6. G. Lim, Tool-wear monitoring in machine turning, J. Materials Processing Technology 51 (1995) 25-36 7. S. Kurada, C. Bradley, A machine vision system for tool wear assessment, Tribology International 30 (1997) 295-304 8. T. Otto, L. Kurik, J. Papste, A digital measuring module for tool wear estimation, DAAAM Intl. Scientific Book (2003) 435-444 9. W. Wang, Y. Wong, G. Hong, Flank wear measurement by successive image analysis, Computers in Industry 56 (2005) 816-830 10. W. Wang, G. Hong, Y. Wong, Flank wear measurement by a threshold independent method with subpixel accuracy, Intl. J. Machine Tools & Manufacture 46 (2) (2006) 199-207 11. Y. Kwon, M. Jeong, O. Omitaomu, Adaptative support vector regression analysis of closed-loop inspection accuracy, Intl. J. Machine Tools & Manufacture 46 (6) (2006) 603-610 12. F. Kim, I. Choi, Development of a tool failure detection system using multisensors, Int. J. Machine Tools & Manufacturing 36 (1996) 861-870 13. R. Silva, R. Reuben, K. Baker, Tool wear monitoring of turning operations by neural network and expert system classification of a feature set generated from multiple sensors, Mechanical Systems and Signal Processing 12 (1998) 319-332 14. D. Choi, W. Kwon, C. Chu, Real-time monitoring of tool fracture in turning using sensor fusion, Int. J. Advanced Manufacturing Technology 15 (1999) 305310 15. N. Ghosh, Y. Ravi, A. Patra, S. Mukhopadhyay, S. Paul, A. Mohanty, A. Chattopadhyay, Estimation of tool wear during CNC milling using neural network-based sensor fusion, Mechanical Systems and Signal Processing (2006) In Press 16 16. M. Shiraishi, Scope of in-process measurement, monitoring and control techniques in machining processes - Part 1: In-process techniques for tools, Precision Engineering 10 (1988) 179-189 17. M. Khan, M. Abou-Zeid, On-line sensing of tool wear for adaptative control of turning operation, SME Technical Paper, Dearbon, Michigan, 1991 18. G. Byrne, D. Dornfeld, I. Inasaki, G. Ketteler, R. Teti, Tool condition monitoring (TCM) - the status of research and industrial applications, Annals of the CIRP 44 (1995) 541-567 19. P. Prickett, C. Johns, An overiew of approaches to end milling tool monitoring, Int. J. Machine Tools & Manufacture 39 (1999) 105-122 20. D. Dimla, Sensor signals for tool-wear monitoring in metal cutting operations - a review of methods, Int. J. Machine Tools & Manufacture 40 (2000) 10731098 21. B. Sick, On-line and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research, Mechanical Systems and Signal Processing 16 (2002) 487-546 22. E. Jantunen, A summary of methods applied to tool condition monitoring in drilling, Int. J. Machine Tools & Manufacture 42 (2002) 997-1010 23. J. Jurkovic, M. Korosec, J. Kopac, New approach in tool wear measuring technique using CCD vision system, Intl. J. Machine Tools & Manufacture 45 (9) (2005) 1023-1030 24. F. Al-Sulaiman, M. Baseer, A. Sheikh, Use of electrical power for online monitoring of tool condition, J. Materials Processing Technology 166 (2005) 364-371 25. T. Pfeifer, L. Wiegers, Reliable tool wear monitoring by optimized image and illumination control in machine vision, Measurement 28 (2000) 209-218 26. M. Lanzetta, A new flexible high-resolution vision sensor for tool condition monitoring, J. Materials Processing Technology 119 (2001) 73-82 27. W. Wang, Y. Wong, G. Hong, 3D measurement of crater wear by phase shifting method, Wear, 261 (2) (2006) 164-171 28. F. Giusti, M. Santochi, G. Tantussi, On-line sensing of flank and crater wear of cutting tools, Annals of the CIRP 36 (1987) 41-44 29. K. Prasad, B. Ramamoorthy, Tool wear evaluation by stereo vision and prediction by artificial neural network, J. Materials Processing Technology 112 (2001) 43-52 30. L.K. Hernández, H. Cáceres, J. Barreiro, E. Alegre, M. Castejón, R.A. Fernández, Monitorización del desgaste de plaquitas de corte usando visión artificial, in: Proc. VII Congreso Iberoamericano de Ingeniería Mecánica, México D.F., 12-14 October, 2005. 31. J. D. Banfield, A. E. Raftery, Model-based Gaussian and non-Gaussian clustering, Biometrics 49 (1993) 803-821 32. C. Fraley, A. E. Raftery, MCLUST: Software for model-based cluster analysis, Journal of Classification 16 (1999) 297-206 33. C. Fraley and A. E. Raftery, Model-based clustering, discriminant analysis, and density estimation, Journal of the American Statistical Association 97 (2002) 611-631 18 34. C. Fraley and A. E. Raftery, MCLUST: Software for model-based clustering, discriminant analysis, and density estimation, Technical Report No. 415, Department of Statistics, University of Washington, 2002. 35. C. Fraley and A. E. Raftery, Enhanced Software for model-based clustering, discriminant analysis, and density estimation: MCLUST, Journal of Classification 20 (2003) 263-286 36. A Language and Environment for Statistical Computing (ISBN 3-900051-070), Development Core Team, Foundation for Statistical Computing, Vienna, Austria, http:/www.R-project.org, 2003 37. A.P. Dempster, N.M. Laird, D.B. Rubin, Maximum likelihood from incomplete data via the EM algorithm, J. Royal Statistical Society 39 (1977) 138. G. McLachlan, T. Krishnan, EM algorithm and extensions, Wiley, New York, USA, 1997 39. G. Schwarz, Estimating the dimension of a model, Annals of Statistics 6 (1978) 461-464 40. R.A. Fisher, The use of multiple measurements in taxonomic problems, Annals of Eugenics 7 (1936) 179-188 41.R.O. Duda, P.E. Hart, D.H. Stork, Pattern Classification (ISBN 0471056693), Wiley Interscience, 2000 42. R. Kohavi, F. Provost, Glossary of terms, J. Machine Learning 30 (2-3) (1998) 271–274 43. E.B. Fowlkes, C.L. Mallows, A method for comparing two hierarchical clusterings, Journal of the American Statistical Association 78 (1983) 553-569. 44. M. Meila, Comparing clusterings, Technical report 418, University of Washington, 2002 20 Figure 1. Equipment used for the tests: vision system and lathe Figure 2. (a) First images in a series showing three wear levels. (b) Segmented images with the worn region in white. Figure 3. LDA projection showing the three clusters as well as the boundaries amongst clusters. Figure 4. LDA projection with wear evolution for an insert Figure 5. Probability of belonging to the dominant wear cluster at each sample Table 1. Confusion matrix for Mclust clustering results vs. LDA class prediction. Table 2. Weights of descriptors in the classification 22 Figure1 Figure2 (a) (b) Figure3 Figure4 Figure5 Table1 L1 L2 L3 M1 426 43 0 M2 16 611 1 M3 2 44 240 Table2 Parameter LD1 LD2 % LD1 % LD2 Eccentricity 10,167570 36,765440 57,55% 74,70% Extent 3,705677 9,383085 20,97% 19,06% Solidity 3,549876 2,664709 20,09% 5,41% EquivDiameter 0,126829 0,237292 0,71% 0,48% Orientation 0,072087 0,071810 0,40% 0,14% MajorAxisLength 0,023715 0,031517 0,13% 0,06% MinorAxisLength 0,018626 0,060774 0,10% 0,12% Area 0,000353 0,000546 0,0019% 0,0011% ConvexArea 0,000353 0,000145 0,0012% 0,0003%