Vision-based Quality Inspection for Components with Small Diameter and Deep Hole

Transcription

1 Vision-based Quality Inspetion for Components with Small Diameter and Deep Hole Shool of Computer and Control Engineering University of Chinese Aademy of Sienes Beijing , China Xinyi Gong, Hu Su, Zhengtao Zhang, and De Xu {gongxinyi2014, hu.su, zhengtao.zhang, Researh Center of Preision Sensing and Control Institute of Automation, Chinese Aademy of Sienes Beijing , China Abstrat - Fousing on the long-thin ylindrial omponent with small diameter and deep hole, we present an internal struture inspetion method based on miro vision. A amera onsisting of CCD sensor and endosope is used to get lear images of the omponent s internal struture. The urves in the image are extrated to represent the internal struture. A searh strategy based on Kalman filter is proposed to detet the internal wire. Then, the quality of the omponent is inspeted aording to the extrated urves. Experimental results show that the proposed searh strategy ould reliably and preisely extrat urves, oupying the apaity of overoming the restritions suh as low-ontrast onditions, emergenies of noises and shadows. Moreover, experiments are well onduted to verify the effetiveness of the proposed methods. Index Terms - long-thin ylindrial omponent, small diameter, deep hole, Kalman filter, urve extration, lassifiation, quality inspetion. I. INTRODUCTION Component with small diameter and deep hole is basi and important in the fields of aerospae, navigation and military industry. The integrity of the internal struture determines its ondutive properties, and thus affets the quality of the omponents. Currently, most of enterprises are using manual detetion method, whih is diffiult to ahieve well repeatability, effiieny and auray performane[1-5]. Automati inspetion of the advaned omponent is of great signifiane to improve the quality of the parts and redue the ost of the produts[6,7]. Faing the problem of inspetion of miro omponent, it is a hard task to ahieve satisfatory auray for industrial appliations. Atually, it is even diffiult to get lear image of the omponent due to the small size and omplex struture. Aording to previous few publiations, the inspetion system whih is based on miro vision has attrated more and more attention. The system ould greatly improve the preision, effiieny and automation [8-10]. For example, a miro vision system was proposed in [11]. However, the system has a small field of view (FOV) and low depth of field (DOF). Using the system, it an hardly obtain lear image of internal struture of the aforementioned omponent. Due to the omplexity and low ontrast of the image, ommon edge detetion algorithm an hardly extrat the internal struture aurately. In addition, the noises and shadows introdue more diffiulties. In [12], Patrauean et. al. proposed a detetor for line segmentation and elliptial ar detetion based on line segmentation detetor (LSD) [13], alled ELSD(Ellipse and Line Segment Detetor), whih formally guarantees the ontrol of the number of false positives and requires no parameter tuning. Instead of using a mass of lines to generate the ar based on LSD, the method an detet the line, the ellipse ar and other urves. However, the generality of the model is limited and it annot be applied for the feature extration in the omponent disussed here. In [14], an edge detetion algorithm based on partile swarm optimization was presented. Compared with other algorithm, it ould extrat all edges in the images and filter out noise more aurately. However, it is far from satisfatory sine it annot distinguish the omponents from shadows. In [15], the extration of different objets on images was viewed as a minimal overing problem from a new perspetive. In this way, different omponents an be well distinguished. But on the other hand, the method annot detet the ontour aurately and determine the qualifiation of the parts well. In this paper, an automati feature detetion method is proposed, whih onsists of image aquisition and the orresponding searh strategy, to effiiently and preisely extrat the urve features of long-thin ylindrial omponents with small size and deep hole. In the image aquisition, we ombine endosope with CCD to get lear images of internal struture of the omponent. Based on Kalman filter, a urve extration algorithm is also proposed to extrat the urves in the image. With the extrated urves, the internal struture of the omponent ould be represented and the quality of the omponent is thus determined. The rest of the paper is organized as follows. In Setion II, we introdue the hardware platform and the omponents to be deteted. The feature extration algorithm of internal wires is disussed in Setion III, whih extrats features of the struture inside the omponents. In Setion IV, the lassifiation of the omponents quality is introdued. In Setion V, experiments are provided to show the effetiveness of the proposed method. Finally, the paper is onluded in Setion VI. II. COMPONENTS AND HARDWARE PLATFORM A. Introdution of omponent Fig. 1 represents the omponent to be inspeted. The diameter of the hole is 1.1mm and the depth is 5.3mm. The diameter of the wires in the hole is 0.12mm. Eah omponent is with 5 opper wires, overing from the top to the bottom.

2 The shapes of the opper wires determine the ondutive property, and thus influene the quality of the omponent. When all the urves representing the wires have been extrated, the quality of the omponent ould then be determined. In this sense, effetive extration of the wire urves is the first onsideration. An iterative searh algorithm is proposed based on Kalman filter. They are irumferene uniform, and the angle of every two adjaent wire is 72 degree, as illustrated in Fig. 1. B. Hardware platform As an be seen in Fig. 2, the hardware platform inludes the following parts: horizontal and vertial shift platforms, CCD, optial soure, image apture ard and endosope. The horizontal platform has three translational degrees of freedom. The vertial one has three rotational degrees of freedom. The horizontal platform adjusts the distane between the endosope and the omponent along the X axis, Y axis, and Z axis. The vertial platform rotates around the X axis, Y axis, and Z axis, making the omponent and optial axis of the endosope be onsistent. Fig. 2 The developed hardware platform, the hardware platform, the relative position of the amera and omponent A. Determination of the searh initial points Sine low ontrast, emergenies of shadows and noises would possibly interfere the extration of internal opper wires, we firstly determine the initial slope of the starting point and the opper wire. A standard omponent is presented in Fig. 1. In the figure, it an be seen that, with the extension from the inner ring to the enter of the omponent, the wire on the inner ring is shown as a large white spot while near the ring it is shown as a blak area. Aording to the analysis above, the determination of the searh point is made up of two steps, whih respetively are the determination of the inner ring and the determination of the opper wire position. The detailed proess is as follows: the first is to extrat the feature points on the ring to form a irle by the least square method (LSM). For eah point on the ring, a small neighborhood is defined and the sum of gray values of the points in the neighborhood is alulated. Representing the point by the polar angle relative to the enter, we get a series of pairs. Eah pair is orresponding to a ertain point on the ring and onsists of the angle of the point and the sum of the gray values in the neighborhood. Consequently, the maximum and minimum values of the sums and their orresponding angles an be determined, whih are regarded Fig. 1 Component piture and sketh map, Component, Sketh To overome the diffiulty aused by the small diameter and deep hole, the optial soure should ooperate with the endosope. Light ould get into the deep hole through the endosope and the quality of the image would thus be improved. The CCD aptures the image of the omponent via endosope and the image ould be obtained with the image apture ard. Fig. 2 shematially explains the relative positions of the omponent and the lens, whih has a signifiant influene on the larity of the image. With a lot of experiments and analyses, we onlude that the distane between the end of the omponent and the lens along the optial axis diretion is about 1mm. And the distane between the oaxial wire hole and the amera lens should be less than 0.6mm. III. SEARCH STRATEGY 2017

3 as the initial searhing points and an be used to ompute the initial slope of the opper wire. B. Searh the next position on the wire As an be seen from the image, the gray values of the points on the wire are lower than that of the others. Some disrete points with minimum gray values are extrated to represent the wire. Aordingly, we searh these points in the diretion perpendiular to the tangent of the wire. Fig. 3 shows an iterative step of searhing the next point aording to the last and urrent points. Intuitively, the next point is roughly in the diretion determined by the urrent and last points. And the andidate of the next point an be obtained. Then a neighborhood is defined, whih is entered in the andidate and the diretion is perpendiular to the searh diretion. The point with minimum gray value in the neighborhood is determined and is believed to be the next point. The measured slope is omputed with the extrated point and the urrent point. Fig. 3 Searh Strategy C. Searh diretion alulation based on Kalman filter For easier explanation, we define some notations before proeeding to the next setion. The estimated values of the slopes in urrent and previous steps are denoted as k and k p, respetively. The urrent measured value of the slope is k d. In the onstrution of Kalman filter, the following notations are referred. The estimated system state at time t is denoted as X t. The ontrol to the system at time t is denoted as U t. The parameters of the system are denoted as A and B. The output of the system is measured as Z t. The transfer matrix between the input and output variables is H. The system noises are W t and V t and their ovariane matrixes are respetively denoted as Q and R. In order to solve the problem of wire shape variety, low ontrast, shadow, noise, and unertainty about the aurate loation of the opper wire, an urve extration algorithm is proposed based on Kalman filter. Kalman filter is the optimal estimation of linear unbiased minimum variane riterion. The mathematial model of Kalman filter an be onstruted as X t = AX t 1 + BUt + W (1) t Z = HX + V (2) t t t Considering the urrent problem, Kalman filter an be utilized to alulate k from k p and k d. Then, the next opper wire position an be found. The reursion formula is as follows. Supposed that no state ontrol is imposed on the system and that the urrent slope is same to the slope of the last state, A and U t ould be regarded as 1 and 0, respetively. And k is estimated as k = k p (3) The predition ovariane matrix P is alulated as P = P Q (4) p + Supposed that parameter of the measurement system H is 1, then the Kalman gain matrix K g (Kalman Gain) is estimated as Update is estimated as K k = P / ( P R) (5) g + = k + K ( k k ) (6) The estimated ovariane matrix is P = ( 1 K ) P (7) IV. QUALITY INSPECTION A. Quality lassifiation With the extrated urves, the internal struture of the omponent ould be represented and its quality is inspeted. In the following onditions, the omponent is judged to be unqualified, (1) wire laking, whih means that one or more wires are missing; (2) wire broken, whih means that one or more wires are broken; (3) wire linking, whih means that at least one wire is onneted to another; (4) wire tortuous, whih means that one or more wires are more ontorted than the standard; (5) wire nonuniform, whih means nonuniform distribution of the wires. Based on the above onditions, an unqualified omponent ould be lassified by the following rules. The omponent is onsidered to be (1) wire laking, if the amount of the extrated urves is less than the standard; (2) wire broken, if one or more urves are shorter than the standard; (3) wire linking, if two or more urves are onneted to eah other; (4) wire tortuous, if the distortion of the urves is bigger than the standard; (5) wire nonuniform, if the distanes between the initial points of two adjaent urves is smaller than the standard; (6) qualified, otherwise. Eventually, the quality of the omponent is inspeted with the rules mentioned above. B. Proedure of the quality inspetion The flowhart of the inspetion algorithm is presented in Fig. 4. Ring extration is firstly onduted. The start point and the initial searhing point for eah wire are determined. The initial slope is alulated with the two points. With the start point and the initial slope, we get the andidate of the next point. Searhing for the gray extreme value point in the viinity of the searhing leads to the next measuring point. p g d g

4 The slope of the urrent point to the next searh point is alulated, and served as the urrent slope measurement. Finally, the atual slope is estimated by using the Kalman filter algorithm, based on whih the next searhing point is determined. The proedure ontinues until the features meet the termination ondition. After extrating all urves in the omponent, we an inspet the quality of the omponent. Start General Searh Ring Extration Extrat Initial Searhing Point Searh Next Point Calulate Searh Diretion Termination Condition Satisfied? N Kalman based Searh Update Slope Update Current oordinate Fig. 5 Experimental results with and without Kalman filter, result with Kalman filter, result without Kalman filter B. Curve extration experiment To verify the effetiveness of the proposed searh strategy, serial experiments were onduted, with one of experiment results was listed as below. In the experiment, irle extration result was shown in Fig. 6. In Fig. 7, the sum of the gray values of points on the ring was alulated to searh the start point on the wire. And then the extreme points were obtained. In Fig. 7, the oordinates of the orresponding points in the image were determined and regarded as the initial wire searh value. The results were displayed as the green irles in Fig. 6. Afterwards, using the searhing strategy, points on the opper wire were searhed one by one, as an be seen in Fig. 8. Fig. 9 showed the final urve extration results. Y Classifiation Finish Inspetion Fig. 4 Algorithm Flow Chart V. EXPERIMENTS AND RESULTS The host omputer used for experiment was Intel Core TM 1.46GHz. The parameters of Kalman filter were set as R=10, Q=R exp(-k 2 /4000), where k is the iterative number. A. Comparison experiments with and without Kalman filter To verify the neessity of introduing Kalman filter in urve extration, omparison experiments were onduted. As demonstrated by Fig. 5, the expeted urves are marked in yellow while the extrated in red. Experimental results with and without Kalman filter are shown in Fig. 5 and Fig. 5 respetively. It an be seen that the urve extrated with Kalman filter oinides the target better and thus an effiiently overome the negative influene aused by noises and shadows, and guarantee the auray of the result of quality inspetion. Fig. 6 Results of intermediate steps Cirle extration result, Copper wire on the ring To further verify the effetiveness of the proposed method, totally 856 experiments were onduted on omponents with different types of defets. The inspetion results were listed in Table I. From Table I, the preision rate of the algorithm was 99.37%; the auray was 99.77%; the reall rate was 99.37%; and the auray of repeat-ability test was 95%. We an onlude from the results that satisfatory auray is ahieved with the proposed algorithm.

5 Fig. 7 Copper wire position and searh start point, the sum of the gray values of points on the ring,? Defets Table I Results statistis Samples Inorretly Classified Error Rate (perentage) Wire Laking Wire Broken Wire Linking Wire Tortuously Wire nonuniform Qualified VI. CONCLUSION This paper presents an automati quality inspetion method of the omponents whih have small diameter and deep hole. In the image aquisition proedure, the endosope is ombined with CCD to obtain lear image of the omponent. In the urve extration proedure, the estimate and the Kalman gain of the slope are dedued based on Kalman filter. And the urves represented by some disrete feature points are extrated by searhing gray extreme point. In addition, rules are onstruted for quality inspetion of the omponent. The experiment results show that, the auray rate of the proposed method reahes 99.7%, higher than the artifiial inspetion auray. Thus, it an be onluded that, requirement on auray for industrial appliations is satisfied, and the effetiveness of the proposed method is verified. Fig. 8 Steps of searhing Fig. 9 Final detetion result ACKNOWLEDGMENT The authors would like to thank the reviewers for their efforts to improve the quality of the paper. This work was supported by the National Natural Siene Foundation of China under Grants , , and And the work was also supported by the Youth Innovation Promotion Assoiation, CAS( ). REFERENCES [1] Z. Y. Xu, J. K. Li and, J. Ling, Fast preision measurement of the apertures of eletri onnetor shell based on halon, Manufaturing Automation, vol. 18, pp , [2] E. Kai. and M. Katsumi, Miro-drilling of monorystalline silion using autting tool, Preision Engineering, vol. 25, no. 3, pp , [3] C. Dunsky, H. Matsumoto, and G. Simenson, High-speed mirovia formation with UV solid-state lasers, in Proeedings of the IEEE, vol. 90, no. 10, pp , [4] B. H. Yan, A. C. Wang, and C. Y. Huang, Study of preision miroholes in borosiliate glass using miro EDM ombined with miro ultrasoni vibration mahining, International Journal of Mahine Tools and Manufature, vol. 42, no. 10, pp , [5] M. Lai, T. Li, and F. Li, Modeling and analysis of miro sale milling onsidering size effet,miro utter edge radius and minimum hip thikness, International Journal of Mahine Tools & Manufature, vol. 48, no. 1, pp. 1-14, [6] Y Zhang, B K Chen, X Liu, Autonomous Roboti Pik-and-Plae of Miroobjets. IEEE Transations on Robotis, vol. 26, no. 1, pp ,2010 [7] Q Xu, Y Li, N Xi. Design, Fabriation, and Visual Servo Control of an XY Parallel Miromanipulator With Piezo-Atuation. IEEE Transations on Automation Siene & Engineering, vol. 6, no. 4, pp , [8] T. Kurita, and M. Hattori, Development of new-onept desk top size mahine tool, International Journal of Mahine Tools & Manufature, vol. 45, no. 7, pp , [9] G. Phillip, and S. G. Kapoor, A new aeleration-based methodology for miro/meso-sale mahine tool performane evaluation, International Journal of Mahine Tools & Manufature, vol. 46, no. 12, pp , [10] S. Rubanovl and P. R. Munroe, The appliation of FIB milling for speimen preparation from rystalline germanium, Miron, vol.35, no. 7, pp , [11] D. Xu, F. Li, Z. Zhang, et al., Charateristi of monoular mirosope vision and its appliation on assembly of miro-pipe and miro-sphere, in Proeeding of 32nd Chinese Control Conferene, pp , [12] V. Patrauean, P. Gurdjos and R. G. Von Gioi A parameterless line segment and elliptial ar detetor with enhaned ellipse fitting, in Proeedings of European Conferene on Computer Vison, pp , [13] R. G. von Gioi, J. Jakubowiz, and J. M. Morel, LSD: A fast line segment detetor with a false detetion ontrol, IEEE Transations on Pattern Analysis and Mahine Intelligene, vol. 32, no. 4, pp , 2010.

6 [14] M. Setayesh, M. Zhang, and M. Johnston, A new homogeneity-based approah to edge detetion using PSO, in Proeedings of IEEE Conferene on Image and Vision Computing, pp , [15] P. Felzenszwalb and D. MAllester, A min-over approah for finding salient urves, in Proeedings of IEEE Conferene on Computer Vision and Pattern Reognition Workshop, pp , 2006.