Optical Pattern Inspection for Flex PCB Challenges & Solution

Proceeings of the 17th Worl Congress The International Feeration of Automatic Control Optical Pattern Inspection for Flex PCB Challenges & Solution Rambabu K, Dinesh Kumar M, Vijay Kumar K, S P Nath*, Roh, Young Jun ** * Tata Elxsi Limite, Whitefiel Roa, Hooy, Bangalore-560048 Inia (Tel: 91-8022979123; e-mail: spnath@tataelxsi.co.in). **LG-PERI, South Korea Abstract: Due to the material properties an technology avancement in printe circuit flex tapes, traitional PCB inspection algorithms using reference image cannot be use. Flex tapes are flexible circuits enabling the esign an prouction of lighter, faster an smaller electronic proucts. A flex tape is mae of flexible polymer film that is laminate onto a thin sheet of conuctive material an etche to prouce a circuit pattern. The with of fine pitch patterns can be of a few microns. Due to the material characteristics of printe circuit tapes, such as thinness an flexibility, the images capture uring inspection are istorte (stretch or shrink, tilt, meanering, etc.). These anomalies an shining tape surface make the possibility of using conventional PCB inspection algorithms using referential methos for pattern inspection, ifficult. The goal of pattern inspection is to ientify pattern efects open, short, nick, protrusion, islan, pinholes an ust particles. Non-repetitive an non-uniform istortions are some of the main challenges in inspecting these flexible tapes. Some of the other challenges face uring the esign of the efect etection algorithm are non-uniform lighting an the gamut of sizes, shapes & orientation of efects as well as features on the printe circuit. Moreover, the algorithms nee to take into account the stringent efect tolerance limits followe by the inustry. 1. INTRODUCTION There are two kins of package substrates rigi substrate an tape substrate. Tape substrate, which is thin an flexible, is often referre to as Flex Tape. It offers superior thermal an electrical performance over thicker substrates. Therefore, flex tapes fin applications ranging from mobile consumer proucts to magneto-resistive hea isk rives. A typical flex tape substrate is a copper polyamie or copper ahesive polyamie structure use as the ielectric material. In packaging methos such as Tape Automate Boning (TAB) an Chip On Film (COF), the chip is mounte on a flexible tape, containing the copper contact lines. The resulting package calle Tape Carrier Package (TCP) provies flexible interconnects. A tape substrate is compose of high-strength an hightemperature polymer material an hence is suitable for mobile applications, printers an isk rives. Their flexibility allows the rea/write hea to move over the isk, thus allowing evelopment of faster an higher-ensity rives. Flex tapes are also foun in flat panel isplays, like LCD (Liqui Crystal Display) an plasma screens. Several isplay river chips nee to be mounte aroun the perimeter of these isplays. As these chips have flexible contacts, the contacts are connecte to the front of the isplay an the chip is then fole an attache to the back of the isplay, using minimal space. Fine patterns in the orer of few micrometers can be achieve on the tapes, thus increasing the component ensity. Flex tape substrates for IC packages use for Tape Ball Gri Array (TBGA) an Chip Scale Package (CSP) take avantage of the fine-pitch wiring possible on flex tapes. Other salient characteristics are reuce weight, space, increase reliability & repeatability proviing uniform electrical characteristics for high-spee circuitry, less power consumption an reuce assembly costs. The paper is organize in multiple sections. Following section provies overview of current PCB inspection algorithms, their merits an emerits. Section 3 iscusses challenges an complexity in algorithm esign for flex tape optical inspection. The paper aresses the approaches to overcome these challenges in section 4; section 5 an 6 gives algorithm approach an flow respectively. Section 7 represents results followe by conclusion an references. 2. PCB INSPECTION ALGORITHMS Printe circuits are inspecte to isolate efects before insertion of components an solering process. The pattern efects that affect the copper patterns on flex circuits inclue open (break in the conuctor), short (copper briges between conuctors), nick (partial open of conuctor), protrusion (copper spurs), islan (spurious metal), pinhole (over etch of conuctor) an ust particles. 978-1-1234-7890-2/08/$20.00 2008 IFAC 8196 10.3182/20080706-5-KR-1001.2847

The current machine vision algorithms use for PCB inspection can be broaly classifie into three categories base on the paper (Moganti 1996) referential approaches, non-referential approaches an hybri approaches as shown in Fig. 1. In referential approach, a test image is compare with the reference image for efect etection. Reference base inspection, in turn, can be ivie into two types. 1. Image comparison techniques: These approaches are base on image comparison, between pixels in the test image an an iealize reference image a. Image subtraction The test image is compare against the image of an ieal sample (logical XOR operation). b. Template matching Comparing a template with an object in an image an it is usually performe at the pixel level as with template correlation with reference to the paper (Maneville 1985). c. Phase-Only metho Stanar template matching technique, base on phase only correlation between test an reference images as referre from the paper (Koichi Ito 2004). 2. Moel-Base inspection methos: These approaches involve recognition of circuit features in the test image followe by a comparison against a set of reference features a. Graph matching metho Base on the topological/structural comparison, which compares the stanar graph, obtaine from the conuctor an insulator image patterns of the test an reference images. b. Tree c. Syntactic The isavantages of referential approaches are that they are sensitive to noise, rotational an translation errors, image alignment errors an image istortion. Besies, these methos are time consuming. In non-referential approaches, the image is transforme into a atabase of the patterns or features present in the image. Design rule verification, generic property verification or feature recognition algorithms are applie for efect etection. Non-referential inspection can be broaly classifie into two methos: 1. Morphology processing These transformations employ specific sequences of neighborhoo transformations to measure useful geometric properties in images as referre from the papers (Louisa Lam 1992, Qin-Zhong Ye 1988). 2. Encoing techniques a. Bounary analysis This metho is base on the representation of the bounaries in a traceable form, followe by a rule verification proceure. b. Run-length encoing Operations such as bi-image processing, pattern recognition, morphological operations, local mask operations, coorinate transformations an feature extraction using Runlength encoing (RLE) moe for ientifying the efects fall into this category. The isavantage of non-referential approaches is that they require the stanarization of feature types an cannot etect efects that o not violate the esign rules. Hybri approaches for PCB inspection make use of both referential an non-referential approaches to overcome their inherent rawbacks. These methos can etect missing features, efects that o not violate esign rules an extraneous features. Hybri approaches can etect efects irrespective of feature size on the printe circuits. Some algorithms base on hybri technique are: 1. Generic metho This metho compares a small list of preicte feature types an locations with a list of etecte features. 2. Pattern etection using Bounary analysis This technique uses a hybri flaw-etection metho base on pattern etection an bounary analysis. 3. Circular pattern matching In this metho, a template comparator is use to perform the encoing, while the 8197

efect etection logic is use to verify the coes to juge if the coes are contraicting. 4. Learning methos a. Raial matching algorithm b. Shape comparison metho 3. CHALLENGES AND COMPLEXITY IN FLEX TAPE INSPECTION Several challenges arise uring the esign of algorithms for efect etection on flex tape circuits ue to the material characteristics of flex tapes such as thinness an flexibility, istortions such as stretch/shrink, tilt an meanering that may appear in the capture images uring optical inspection. Distortions can also appear in the flex tapes ue to mechanical feeing errors uring inspection of these flex tapes, thus increasing the complexity involve in the esign of approaches for inspection. Since the image istortion is non-uniform an non-repeatable, conventional PCB inspection algorithms using referential approaches cannot be use for inspection of bare printe circuits on flex tapes. The shining surfaces of flex tape results in non-uniform contrast of the capture image. Thus inspection algorithms will not be able to process the complete image at a time for efect etection. Defects such as open an shorts of conuctor traces ientifie in the test image nees to be verifie using a reference image. Hence test an reference images shoul be aligne for translation an rotational errors. These translation an rotational errors are localize errors an are non-uniform an non-repetitive. Flex tapes enable circuits to have features with size in the orer of micrometers. Typically, the smallest pattern efect size shoul be at least twice the size of the vision system resolution. In other cases, sub-pixel accuracy in the ege etection, imension measurements, an alignment stages is require for ientification of efects on fine pitch pattern regions. The price/performance of the computation power require to hanle the extra ata must be taken into consieration uring the esign of the inspection system. Defects on printe circuits can be of any size an shape. They can be wie shallow, wie eep, narrow shallow or narrow eep as shown in Fig. 2. example, the algorithm base on the Eigen value of covariance matrix as referre from paper (C Yeh CH 2001) fails in some of the above cases. Similarly, patterns (features) on the flex tapes can have varying size an orientation. Patterns can be horizontal, vertical or of any slope or a combination of these. Inspection algorithms shoul be able to ientify the efects on any of such patterns an in corner areas. The valiation of a possible efect as an actual efect involves the complexity of comparing the caniate efect s with with respect to pattern an/or space with. If the efect with is greater than a factor of corresponing pattern an/or space with, then it is valiate as a efect. Image pre-processing like binarization of the image an ege etection is require prior to efect etection. Since the image contrast is non-uniform an non-repetitive, binarization as referre from the papers (Ping-Sung Liao 2001, Mehmet Sezgin 2004, Sungzoon Cho 1989) is a challenge in the algorithm esign. The ege-etection algorithm shoul yiel single pixel with eges without ege breaks. Existing graient-base methos give ege breaks in weak pixel graient pattern areas. Moreover, contour retrieval algorithms expan the patterns an hence moify the efects. Designing an appropriate ege etection algorithm is another challenge in the inspection algorithm esign. 4. APPROACHES TO OVERCOME THE CHALLENGES 4.1 Image istortion Due to the physical characteristics of flex tapes, translation an rotational errors may appear in the image capture uring inspection. Since these istortions are non-uniform an nonrepetitive, the image woul nee to be ivie into small frames an must be aligne to the corresponing reference frames. Techniques for aligning test an reference images are: a) Cross Correlation (Phase only correlation): Fourier transforms of test an reference image are compute as F an G using Fast Fourier Transform (FFT). The cross-phase spectrum (or Normalize cross spectrum) efine as: R FG = F * G / F * G, (1) Where G is the complex conjugate of G as prove in the papers (Koichi Ito 2004, Barbara Zitova 2003). Next, the inverse transform of R FG (r FG ) is calculate. The position of the highest peak in r FG gives the translation shift between test (F) an reference images (G). Ientifying the esign an classification rules for efect etection, inepenent of the size an shape of efects is a major challenge in the esign of inspection algorithms. For b) Pel Difference Classification (PDC) c) Mean Absolute Difference of pixel values ) Mean Square Difference of pixel values 8198

e) Integral Projection 4.2 Non-uniform contrast levels in image To overcome the non-uniformity of contrast levels in the image, the capture test image is ivie into sub-images an each frame is processe for efect etection. The actual processing winow size is set greater than the sub-image frame size so that efects on bounaries of the frames are also etecte. The overlapping with of the sub-images is set to be equal to the sum of the maximum eviation between test an reference image an the minimum local neighborhoo winow size require for efect etection analysis. Consiering small frames an processing each frame iniviually solves the non-uniform image contrast problem. 4.3 Sub-pixel analysis for efects in Fine Pitch Patterns (FPP) In fine pitch pattern regions of the image, sub-pixel analysis is require for ientifying efects. Approaches that increase resolution of images, or sub-pixeling techniques for ege etection, measurement an alignment can be use for the analysis of efects in such regions. Toay s vision systems have respone by being esigne aroun higher resolution imagers, an have greater computing power to achieve high performance in spite of the large amount of ata to be processe. 4.4 Different sizes an shapes of efects PCB inspection algorithms shoul be able to etect efects inepenent of their size an shape. Defect etection algorithms base on statistical methos fail in ientifying efects inepenent of their size an shape. Ientification of efects shoul be base on local neighborhoo geometry in the ege image. Defect ientification esign rules an classification rules base on bounary analysis techniques that analyze the local geometry can etect efects irrespective of efect size an shape. 4.5 Dust particles ientification Foreign materials may present on flex tapes uring inspection. These foreign materials are calle ust particles, shown in Fig. 3. Dust particles can be ientifie an classifie base on their irregular geometry an pixel color values. 4.6 Consieration of efect tolerance limit The efect fining capability of the inspection system is evaluate by the efect tolerance limits consiere uring the efect analysis phase. The efect with in a irection perpenicular to the pattern is compare with a fraction of the corresponing pattern an/ or space with for efect valiation. If the efect with is greater than a factor of corresponing pattern an/or space with, it coul cause current carrying capacity problems, capacitive effects, electromagnetic effects, etc., that affect the functionality of the circuit 5. DEFECT DETECTION ALGORITHM APPROACH This approach is a hybri approach, which makes use of both non-referential metho an referential metho that are complement to each other leaing to high error sensitivity, irrespective of the feature an efect sizes on flex tapes. Segmentation is use to ientify regions having textual information on the flex tapes. Segmentation also ientifies the fine pitch pattern regions on the flex tapes an separate esign rule methos are applie to them, giving goo accuracy in etecting efects on fine pitch pattern regions. Ientifying efect caniates Rejection/ selection of efects Referential STAGE I Classification STAGE II Mousbite, Protrusion, Islan & Pinhole Nonreferential Design Rule Checking (DRC) Dust particles Open & Short F I N A L D E F E C T S 8199 There are two stages in this approach as shown in Fig. 4, Stage I is a non-referential approach an Stage II is a hybri approach. Stage I: In this stage, non-referential metho ientifies all the efect suspicions for pattern efects. Tracking is one on either sie of each efect suspicion to ientify the actual

efect caniates. These efect caniates are classifie as Open, Short, Mouse-bites, Protrusions, Islans, Pinholes an Dust particles. Stage II: In this stage, referential metho is use to reject all obvious non-efects among the efect caniates ientifie by non-referential metho. For example, open an short efect caniates ientifie by non-referential metho have to be valiate by the referential metho. Design Rule Checking valiates these classifie efects by applying corresponing esign rules an the valiate efects are ae to the final efects list. Design Rule Checking employs ifferent methos for efects on fine pitch pattern regions an coarse pitch pattern regions. 6. DEFECT DETECTION ALGORITHM FLOW Defect etection algorithm for pattern inspection of flex tapes is illustrate in Fig. 5. Test image Segmentation b) Test image frames are selecte with overlapping regions for efect etection as per (2). Effective image size for efect etection = Test image frame size (Minimum with require for tracking aroun the efect + Maximum eviation between test an reference images) (2) c) Segmentation is applie on each test image frame to classify it as fine or coarse pitch region an to ientify textual regions in the image ) Bounaries that have single pixel with an ege continuity are ientifie in the test image frames 6.2 Ientifying efect suspicions Ege image obtaine in the above step is traverse an at each pixel a region of size 3X3 with current pixel as center, is compare with pre-efine templates for ientifying efect suspicions. The stanar templates use are as shown in Fig. 6. Ege etection Classifie as fine or coarse pitch regions Ege image Template matching Defect suspicions Tracking aroun each efect suspicion Fig. 6. Templates use for template matching Defect caniates Defect classification Classifie efects Rejection/Selection Final efects Dust particles Open & Short 6.3 Ientifying efect caniates from efect suspicions a) Bounaries aroun each efect suspicion are tracke at pixel level an the pixel movement irections are note on either sie of the efect suspicion. The following 8 pixel movement irections are consiere as shown in Fig. 7. Reference image Defects to be valiate Design Rule Check (DRC) Mouse bite, Protrusion & Islans/ Pin-holes 135º 90º 45º 180º POI 0º Fig. 5. Defect etection algorithm flow 6.1 Pre-processing a) Since the complete test image of size typically 8000 X 8000 pixels (gray scale) can have fine an course pitch patterns varying from region to region, it is ivie into small image frames of size 256 X 256 pixels for efect etection 225º 270º 315º Fig. 7. Bounary tracking irections b) These pixel movement irections are analyze to verify the efect suspicions as efect caniates 8200

6.4 Defect classification a) Defect caniates ientifie in the above step are classifie base on the following parameters: i. Tracke pixel movement irections aroun the efect ii. Pixel gray value of the mi point of the efect s extreme points b) In the above step, efect caniates are classifie as Open, Short, Mouse bite, Protrusion, Islan, Pinholes an Dust particles 6.5 Classifying Open an Short efects as final efects using reference image Since there can be patterns of Open or Short efect type present in the original image, these efects nees to be verifie by checking for the presence of similar features in the reference image using tracking metho. All Open an Short efects are classifie as final efects at this stage. c) Design rules for Islan an Pinhole efects are: Typical Pinhole efects are shown in Fig. 10. The esign rule for pinhole efect valiation as given in (5). < 1/3 * Pw (5) where, = Defect with Pw = Pattern with 6.6 Classifying Mouse bite, Protrusion, Islan an Pinhole efects as final efects by DRC Pw a) Design rules are use for classifying Mouse bite, Protrusion, Islan an Pinhole efects as final efects base on efect with with respect to corresponing pattern an/or space with b) Design rules for Mouse bite an Protrusion efects are: Typical one-sie an two-sie mouse bite efects are shown in Fig. 8. The esign rule for mouse bite efect valiation as given in (3). > 2/3 * Pw (3) where, = Effective pattern with Pw = Pattern with Fig. 10. Design Rule Check for pinhole efect Typical Islan efects are shown in Fig. 11. The esign rule for islan efect valiation as given in (6) < 1/3 * Sw (6) where, = Defect with Pw = Pattern with Sw Fig. 11. Design Rule Check for islan efect Typical one-sie an two-sie Protrusion efects are shown in Fig. 9. The esign rule for protrusion efect valiation as given in (4). > 2/3 * Sw (4) where, = Effective space with Sw = Space with 6.7 Classifying Dust particles Finally, efect caniates are classifie as ust particles by checking the pixel gray value ranges an irregular geometry. 7. RESULTS Following image shows ientifie efects with above explaine approach. The classifie efects are shown with ifferent color rectangles as shown in Fig. 12. 8201

conventional algorithms. Several proceures to tackle these obstacles have been iscusse in this paper. Fig. 12. Defect Ientification an Classification Fig. 13 shows ientification of efects in special cases like wie-shallow nick, both-sie nick an large short efects with ifferent shapes an sizes Fig. 13. (a) Large short, (b) Shallow wie nick, (c) Large open, () Both-sie nick an (e) Different efect shapes This approach able to yiel results more than 90% efect sensitivity in flex PCB with resolution approximately 8K X 6K, minimum pattern with of 10µm. 8. CONCLUSION Conventional PCB inspection algorithms base on referential approaches o not work in inspection of flex tapes, ue to their inherent material properties. Thus a hybri inspection approach using both referential an non-referential approaches coul be use for pattern inspection of the printe circuits on flex tapes. Statistical metho base algorithms cannot be generalize for etection of all efects inepenent of their size an shape an patterns size on the printe circuits. Bounary technique base local neighborhoo geometry analysis algorithms can etect efects of any size an shape on patterns of any size an shape. Thus the esign of a complete inspection algorithm for the efect etection of printe circuits on flex-tape requires extensive research an evelopment efforts to overcome these challenges face by REFERENCES M. Moganti, F. Ercal, C. H. Dagli an Shou Tsunekawa (March 1996). Automatic PCB Inspection Algorithms: A Survey. Computer Vision an Image Unerstaning, Vol. 63, No. 2, pp. 287-313. J. R.Maneville (January 1985). Novel metho for analysis of printe circuit images. IBM J. RES.Develop, Vol.29, No.1, pp. 73-86. Louisa Lam, Seong-Whan Lee, Ching Y. Suen (September 1992). Thinning Methoologies-A Comprehensive Survey. IEEE Transactions on Pattern Analysis an Machine Intelligence, v.14 n.9, p.869-885. Qin-Zhong Ye, an Per E. Danielson (September 1988). Inspection of Printe Circuit Boars by Connectivity Preserving Shrinking. IEEE Transactions on Pattern Analysis an Machine Intelligence, Vol.PAMI-10, No.5, pp.737-742. Koichi Ito, Hiroshi Nakajima, Koji Kobayashi, Takafumi Aoki, Tatsuo Higuchi (March 2004). A Fingerprint Matching Algorithm Using Phase-Only-Correlation. IEICE Trans. Funamentals, Vol. E87-A, No.3, pp. 682-691. Barbara Zitova, Jan Flusser (2003). Image registration methos: a survey. Image an Vision Computing, 21, pp. 977 1000. Ping-Sung Liao, Tse-Sheng Chen, Pau-Choo Chung (2001). A Fast Algorithm for Multilevel Thresholing. Journal Of Information Science an Engineering, 17, pp. 713-727. Mehmet Sezgin, Bu lent Sankur (January 2004). Survey over image thresholing techniques an quantitative performance evaluation. Journal of Electronic Imaging 13(1), pp. 146 165. Sungzoon Cho, Robert Haralickt, Seungku Yi (1989). Improvement of Kittler an Illingsworth s minimum error thresholing. Pattern Recognition, Vol. 22, No. 5, pp. 609 617. C Yeh CH, Tsai DM (2001). A rotation-invariant an nonreferential approach for Ball Gri Array (BGA) substrate conuct paths inspection. International Journal of Avance Manufacturing Technology. 17:412-24. 8202