Defect Detection and Classification in Ceramic Plates Using Machine Vision and Naïve Bayes Classifier for Computer Aided Manufacturing

Transcription

1 Defet Detetion and Classifiation in Cerami Plates Using Mahine Vision and Naïve Bayes Classifier for Computer Aided Manufaturing 1 Harpreet Singh, 2 Kulwinderpal Singh, 1 Researh Student, 2 Assistant Professor, 1,2 Department of Mehanial Engineering, 1,2 North West Institute of Engineering & Tehnology, Dhudike, Moga, India Abstrat: Detetion of defet is an inevitable task in erami manufaturing. In spite of the importane of the pitorial inspetion of produts, many of the inspetion proesses are performed manually. Here, the development of automati proesses for the examination of erami produts is important. The diffiulty is that manual inspetion has some obvious drawbaks, suh as being long, the high prie involved, and the absene of standardization. Pakaging glass or hina lay produt uses the pakaging ontainers whih of exat size that of the produt to fit in. In this work, we propose a smart system for the automati detetion of the types of defets in erami produts: the previous one is detetion of a defet in erami plates for domesti use, alled erami deformation, and the latter one is lassifiation of a defet alled deformation in plates for domesti use. To evaluate these appliations, we used the dataset made available by the previous researhers for defet detetion purpose only. The result in the form of graphs and GUI are provided showing the effetiveness of the method. In the result it is found that the overall auray for defet detetion in ANN is nearly 73% while this auray for Naïve Bayes is nearly 85%. Also time taken for total defets analysis by Naïve Bayes is almost 0.02s while for ANN it as almost 0.25s. Keywords: Mahine Vision, Inspetion, Defets, Classifiation, Gray Level, Image Series. I. INTRODUCTION Quality is pivotal to the suessful funtioning of an enterprise as it an be explained as the ability of the produts or servies to satisfy the requirement of the ustomers, rather than just onforming to the standards. Quality naive bayes lassifier, quality ontrol and quality improvement are suggested by Juran et al. (1998) as the trilogy whih ensures quality in the modern day manufaturing entres. For an effetive deision at all these stages (in line with Garbage In Garbage Out-GIGO- onept), there needs to be a measure of the urrent status of the quality in an aurate and quik way. A viable option is to use sensors for suh measurements of the ritial parameters of the finished produt. There is another important deision to be made as to the degree of inspetion that is required for any produt. Given the situation of progressive improvement of the manufaturing proess, the inspetion has to be flexible enough to aommodate newer quality standards. When the proess apability is to be improved due to sudden hanges in the quality speifiation, there is a requirement for 100% inspetion. 100% inspetion has its disadvantages, sine the ost involved is higher and additionally the human operators also have to be eduated to make the right deision. But the availability of inexpensive sensors, higher omputer memory storage apabilities and alulation speeds makes it eonomial to implement automated inspetion methodology in the urrent prodution environment. Though sensors are available for automating many of the inspetion tasks, there are still many situations where manual inspetion is the best possible solution. When it omes to visual inspetion of varied kinds of defets in omplex shapes whih are sometimes hidden to the view ompletely, manual inspetion still may be onsidered as a suitable option. Aounting for all shortomings, it has been found that manual inspetion has only 80% auray in making orret inspetions (Juran et al., 1994). As a result there is always an attempt to augment the senses of the inspetor using instruments like sound amplifier, optial magnifier et. The development of an appliation speifi inspetion system is the ultimate form of this. Manual inspetion an be replaed by automati inspetion to redue labour fatigue, and save money and valuable man hours wherever possible. This helps in quality improvement, ost redution, inreased volume, and shorter yle time. Currently in industries, inspetion of omponents is automated to a very large extent. Mahine vision is a form of non-ontat inspetion, whih an be used for a wide range of eletro optial sensing tehniques from triangulation, profiling, 3D objet reognition to bin piking based on image proessing routines. In addition, it is being used for simple detetion and measuring tasks to roboti ontrol. The most ommon appliations are measuring ritial dimensions, deteting flaws, ounting/sorting, assembly verifiation, position, analysis, harater or bar-ode reading/verifiation, and determination of presene/absene of features on small parts. The major advantage is avoidane of human error. The proess involves apturing an image of the objet to be inspeted using vision sensors and then proessing it using image proessing algorithms to measure the features of interest. One very important advantage of mahine vision systems is that 100% inspetion an be ahieved. The mahine vision approah to solve optial inspetion problems an be done in three stages. The first is to set up a amera and proper illumination to IJEDR International Journal of Engineering Development and Researh ( 1195

2 apture images with very good quality required for any speifi appliation. Generally the usage of vision inspetion is ditated by the appliation and it should be always aimed to make it simple. It is more ost effetive and better engineering pratie to apture a good image at the soure, rather than to pour resoures into leaning it up or simplifying it later. The seond stage is the pre proessing of images to gather the neessary quality for the appliation desired and to obtain the image data. Later, the data has to be proessed and then the deision regarding the inspetion status has to be made. The available information has to be used to make a deision regarding the type of defets, sine no approah aimed at defet detetion will be omplete without quantifying the deision. The final step is an elaborate proess in itself beause deision making is a ase of pattern reognition and involves a lot of preliminary steps to ensure orret deisions. II. RELATED WORK Rosati et al. (2009), have attempted real time defet detetion on urved refletive surfaes using a set of mirrors. A speial mirror was designed to diret the illumination at speifi diretion and the presene of defets ould be observed by analysing the images using standard image proessing software. An effiient amera system approah was used, where the mahine vision sensor was in the path of speularly refleted light rays. A similar set up an be used for inspetion of automobile parts with highly refletive surfaes. But the high intensity refletions might saturate the mahine vision sensor resulting in loss of information and this was not disussed in detail. To avoid saturation, redution in the intensity of inident light might ause lesser ontrast in the defets that are imaged. This is a partiular anomaly while using bright field imaging systems. Mirostruture surfaes were reonstruted from SEM images by Samak et al. (2007). Unlike the previously disussed approah, the sample to be inspeted was tilted and the hange in position of the point under observation was onverted into height information using the standard triangulation algorithm used for most of the depth detetion problems. Later De launey s triangulation has been done using this information. The images were mirosopi and the CAD model of many broken materials was suessfully reated suh that frature analysis an be done using this information. Abrahamo vih et al. (2005) desribed the alibration and use of multiple ameras for inspetion. Quite interestingly, the ameras used were ommerially available webams, array of ameras suh that imaging of a large area an be overed. Many of the inspetion senarios like wooden sheet inspetion, sheet metal inspetion demand multiple ameras and the performane, as stated by the author, is really onvining for an industrial user sine the web amera osts 1/50th of an industrial amera. The methodology of imaging different areas of the same surfae and then stithing all of it together will enhane the speed of inspetion provided the illumination soure for all the surfaes also remains the same. But the approah is really benefiial when the surfae for inspetion is planar (whih the author has speifially stated) and an be imaged at one Sun et al. (2005) applied X-ray imaging for real time detetion of welding defets in steel tubes. Though X-ray was used, the signals were aptured in digital form just like visible light and the image was used for subsequent analysis. The use of X-rays and image proessing paradigms ould easily automate the inspetion proess whih was undertaken by human inspetors till then. The feed bak after the implementation of the set up was that it is more effiient than manual inspetion of video streams. This is an example of apturing higher ontrast image information by use of X-rays. Moreover, in this partiular ase, hoies for inspetion set ups were limited sine the defets were subsurfae. Luo and Liou (1998) used multiple ameras to measure the surfae onditions of a ylinder and used the images to measure deviations. The methodology was based on binoular vision. The images aptured by the two imaging devies have been orrelated with eah other to determine the diameter of the sample to be inspeted. Both the imaging devies were used to apture the same surfae under inspetion. Use of two ameras helped in the evaluation of depth information, though there were anomalies due to the urvature of the surfae. The explanation is partiularly useful for determining the diameter of ylindrial omponents, rather than surfae inspetion. Lee et al. (2000) suggested mahine vision system for urved surfae inspetion. Cerami plates mahine are the spindles and plates whih are used in large numbers in eah of the mahines. The resoures available to inspet defets on flat surfaes and irular shapes, none of them are useful for inspeting internal, urved and irular shapes. As a result, thousands of plates are manufatured at the industry to ater to the needs of the onsumer and for the maintenane of older ones. Also, a less omputationally intensive algorithm is required to speed up the entire inspetion proess. Juran et al. (1998) suggested the quality naive bayes lassifier, quality ontrol and quality improvement as the trilogy whih ensures quality in the modern day manufaturing entres. Aounting for all shortomings, it has been found that manual inspetion has only 80% auray in making orret inspetion. A viable option is to use sensors for suh measurements of the ritial parameters of the finished produt. For an effetive deision at all these stages (in line with Garbage In Garbage Out-GIGO- onept), there needs to be a measure of the urrent status of the quality in an aurate and quik way. As a result there is always an attempt to augment the senses of the inspetor using instruments like sound amplifier, optial magnifier et. Luo P.F et al. (1998) disussed the measurement of urved surfaes by stereo vision and error analysis. The omplete vision system has to be designed based on the inspetion requirement whih vary for different appliations. Different methodologies tried out with regards to the kind of illumination will be disussed in this paper. The appliation disussed here in this work is prinipally to detet the presene of defets and lassify them. In many of the industrial appliations the size, dimension, pattern or shape of the objets are assessed. Iivarinen et al. (1998) explained the seletion of features for lassifiation from the segmented images of defets and used Gray Level Co-ourrene Matrix (GLCM) and the related parameters from it for giving input to the lassifier. A.MIvor (2013) in this researh paper bakground subtration model an be used for deteting multiple defets. There are various filters that an be used for subtration of bakground in real time defet detetion i.e. mean shift and kalman filters. Thresholding tehnique is the widest used tehnique for extrating the bakground and foreground of the image. In this method unneessary details an be removed from that image and we an easily extrat the required things from a IJEDR International Journal of Engineering Development and Researh ( 1196

3 real time image. Algorithms analyze the moving objets frames and output of the loation of the objet within that video frame. In early pre proessing stage temporal or speial smoothening is used to eliminate various noises present in the video frame or the image under onsideration. There are various environmental hanges so we also use the tehnique of bakground modeling against those environmental hanges ourring in the video frame. III. PROPOSED METHOD Image proessing methodology to be adopted generally is based on the type of illumination, the harateristis of objet under inspetion, the data to be extrated and type of mahine vision system. Though some of the image proessing proedures are adaptable, most of the new appliations require newer and different tehniques of pre-proessing of the images. The omplete vision system has to be designed based on the inspetion requirement whih vary for different appliations. In many of the industrial appliations the size, dimension, pattern or shape of the objets are assessed. The appliation disussed here in this work is prinipally to detet the presene of defets and lassify them. Different methodologies tried out with regards to the kind of illumination will be disussed in this hapter. Sine speed and auray are the major requirements, eah of the methods will be analyzed for suitability, aording to the speed and the auray at whih the inspetion proess an be ompleted. In our proposed system, the CHT aims to find the irles in the plate image. After generating the edge image, we will onsider the edge points as x k, y ), k=1, 2, 3 n. ( k The weights in a irular Hough spae are analyzed for estimating the three parameters of one irle. A Hough spae an be written as: H ( x, y, r) n k1 h( x k, y k, x, y, r) and k=1, 2, 3 n are the edge pixels. The position with the maximum value of is hosen as the parameter vetor for the strongest irular boundary. However, in our proposed system, CHT is used for deteting the plate outer boundary. Depending upon the database onsidered the defined range of radius searhes for the values that have been set manually. For the given database, a value of the radius for plate ranges between 100 to 110 pixels. In order to formulate the irle detetion proess in an effiient, effetive and aurate naive bayes lassifier, the Cirular Hough transform based approah for the plate was performed first; after the ompletion of this partiular proess, 3 parameters are stored for irles whih inludes, radius and enter oordinates. IV. RESULT The defet lassifiation requires more information about the defets, if a better lassifiation is to be ahieved. The usage of different illumination systems gives more information about the area under onsideration. From the images in Figure, it an be understood that the details are varied in eah of the images, though the same details are being imaged in all of them. Eah of the features has varying highlights in different images. The usage of the window has made sure that eah pixel represents information whih is related to the neighbouring pixels, rather than an abnormal value whih is seldom orrelated to any other pixel values in the viinity. The omparison of the same pixel values in different images will yield a more onrete idea about the presene of a partiular feature in image. The images in Figure prove this aspet. Throughout the implementation of proposed approah, a fusion of information has been used to make better deisions. One major disadvantage of this approah is the number of alulations whih in turn inreased run time for eah algorithm implementation. As a result, it was attempted to onsider a region of interest by reating a mask based on the amera position whih might redue the omputational load and run time by many times. It also helped to simplify the thresholding operation. The thresholding operation was done by finding the optimum ut off frequeny for the parameters under onsideration. An interesting fat to be noted is that only the spatial oordinates of the image is the same at all stages of lassifiation operations, sine the gray sale values have been replaed by statistial parameters whih have their own unique range. As a result, they had to be normalized to a unique sale while doing operations in unison with another pixel. Similar onversion of sale happens in all the measurement proesses in metrology. As a result, the parameters have to be aompanied with a saling fator while implementing the final lassifiation proess also. The images in Figure have the defetive area separated out to a very good extent. Here again, the information fusion has been applied through logial operations to arrive at a final image with an optimum separation of defetive region. The usage of information fusion will help in making deisions on the basis of a large amount of valid information, thus dereasing the possibility of error in the deision. This an be done by avoiding all information, whih is obviously IJEDR International Journal of Engineering Development and Researh ( 1197

4 erroneous, by implementing a ondition for evaluating the suitability of partiular information for use in the algorithm (For example, in a thresholding operation the whole image naive bayes lassifier be showing defetive area of the plate). The parameters in the index values orresponding to the thresholded region is now passed to the defet lassifier. The Naive Bayes Classifier based on normal distribution has been onsidered. All the data points ould not be fitted exatly into the normal distribution. Again, an information fusion approah was onsidered where the data from different ombinations ould be onsidered together and then the best possible deision taken. Fig.4.1: loading of test produt for detetion Fig. 4.2: edge detetion of test produt Fig. 4.3: Image representing irle detetion of test produt IJEDR International Journal of Engineering Development and Researh ( 1198

5 Fig. 4.4: Confusion matrix for ANN Fig. 4.5: Confusion matrix for Naïve Bayes Fig. 4.6: Auray omparison graph for both methods Fig. 4.7: Time omparison graph for both methods The parameters were onsidered two at a time, sine the lassifier beomes erroneous when the ovariane matrix approahes zero (singularity). Instead of using a pseudo inverse whih ould not give onsistent results, inversion has been done in a manual fashion and then the determinant multipliation done at a later stage. IJEDR International Journal of Engineering Development and Researh ( 1199

6 The results are good exept for the ase of yellow stain, whih is a diffiult kind of defet ompared with the rest. Nevertheless many of the pixels orresponding to yellow stain were lassified orretly. The flow hart for the proposed approah is given in Figure 6.16 V. CONCLUSION An approah to inspet and lassify defetive glass plate using mahine vision has been presented in this work. The glass plates were imaged using an experimental set up with a mahine vision system using a speifi system. These approahes may be used to find the presene of defets and suessful lassifiation of defets with very high auray and reliability. Based on this work, the following are the speifi onlusions: 1. Experimental study onludes that plate s defets are lassified into three ategories and these ategories depend upon the type of defets present in plate. In ategory 1, plates have minor defets (like variation in radius of the plate as ompared to the standard plate), ategory 2 plates have large defets (like plates are rak from one side or it is totally broken) and it is send to waste material. Plates whih have only edge defetive are lassified into final ategory i.e. ategory After omparing two methods, study resulted that Naïve Bayes is more aurate and less time onsuming as ompared to ANN. It is also found that auray level for Naïve Bayes is nearly about 85% while for ANN it is nearly about 73%. Also the time taken for total defets analysis by Naïve Bayes is almost 0.02s while for ANN it as almost 0.25s. Table 5.1: Comparison between two methods Name of Method Auray Level Time Taken (Seond) Naïve Bayes 85% 0.02 ANN 73% 0.25 SCOPE FOR FUTURE WORK The proposed set up ould be pratially implemented and fine-tuned to meet the requirement of the industry. Also, instead of the normal halogen light soure, other light soures ould be tried out. There an also be quantifiation of the surfae or geometrial features of the plate by methods suh as phase shifting interferometry and strutured lighting soures et. Suh approahes an be as well used to find the surfae finish of the omponents under onsideration. REFERENCES [1] Rosati G., Boshetti G., Biondi A. and Rossi A. (2009) Real-time defet detetion on highly refletive urved surfaes. Optis and Lasers in Engineering, 47(3-4), [2] Samak D, Fisher A and Rittel D. (2007) 3D Reonstrution and visualization of mirostruture surfaes from 2D images. Naive bayes lassifierals CIRP, 56(1), [3] Abrahamovih G., Barhak J. and Spier P., (2005), Reonfigurable array for mahine vision inspetion (RAMVI), Conferene on Reonfigurable Manufaturing. Naive bayes lassifier Arbor, USA, R-01, May 10 12, [4] Sun Y., Bai P., Sun H.Y. and Zhou P. (2005) Real time automati defet detetion of weld defets in steel pipe. NDT&E International, 38, [5] Luo P.F. and Liou S.S. (1998) Measurement of urved surfaes by stereo vision and error analysis. Optis and. Lasers Engineering, 30, [6] Lee M. F. R., de Silva C. W., Croft E. A. and Wu Q. M. J. (2000) Mahine vision system for urved surfae inspetion. Mahine Vision Appliations, 12, [7] Lindner C., Shäffler F. and León F. P. (2007) Texture-based surfae segmentation using seond-order statistis of illumination series. Leture notes in informatik, 109, [8] A.MIvor, An ative ontour traking method by mathing forground and bakground simultaneously,2013. [9] Iivarinen J. and Visa A. (1998) An adaptive texture and shape based defet lassifiation. 14th International Conferene on Pattern Reognition (ICPR'98), 1, 117. [10] Juran J.M. and Godfrey A.B. Juran's Quality Control Handbook. Fifth Edition, MGraw-Hill. New York, 1998 IJEDR International Journal of Engineering Development and Researh ( 1200