跳到主要內容

臺灣博碩士論文加值系統

(35.172.136.29) 您好!臺灣時間:2021/08/02 03:49
字體大小: 字級放大   字級縮小   預設字形  
回查詢結果 :::

詳目顯示

我願授權國圖
: 
twitterline
研究生:趙寶華
研究生(外文):Victor Ham Choi
論文名稱:Fastandnoise-tolerableimagesubtractionmethodsfordefectdetectiononPCBsandLEDwafers
論文名稱(外文):Fast and noise-tolerable image subtraction methods for defect detection on PCBs and LED wafers
指導教授:蔡篤銘蔡篤銘引用關係
學位類別:碩士
校院名稱:元智大學
系所名稱:工業工程與管理學系
學門:工程學門
學類:工業工程學類
論文種類:學術論文
論文出版年:2009
畢業學年度:97
語文別:英文
論文頁數:129
中文關鍵詞:Automated optical inspectionDefect detectionTemplate matchingImage subtractionStatistical process controlKernelAssembled PCBsLED wafers
外文關鍵詞:Automated optical inspectionDefect detectionTemplate matchingImage subtractionStatistical process controlKernelAssembled PCBsLED wafers
相關次數:
  • 被引用被引用:0
  • 點閱點閱:133
  • 評分評分:
  • 下載下載:0
  • 收藏至我的研究室書目清單書目收藏:0
Automated inspection of assembled printed circuit boards (PCB) is a requirement to ensure the quality of the product and to decrease manufacturing cost. An assembled PCB comprises complex conductive paths with different electronic components such as integrated circuits (IC), resistors and capacitors, which make template matching the only possible approach for automated optical inspection. Current template-based defect detection techniques use simple image subtraction to identify defects on assembled PCBs which relies on accurate image alignment of the reference and inspection images. If this cannot be achieved, noise points are generated around the edges. Changes in illumination are another factor to consider since image difference will yield false detections.
This study is aimed to develop three template-based approaches for defect detection and, especially, focuses on PCBs and light emitting diode (LED) chips. These methods are robust under small misalignments, changes in illumination and manufacturing variation. The first proposed method, named statistical process control (SPC) with multiple templates, uses statistical process control techniques. It takes a number of defect-free images to form a “mean” template, described by the gray-level mean and standard deviation of each individual pixel in the image. This method is faster than the simple image subtraction method, and is tolerant to noise due to misalignments, illumination changes and manufacturing product variation. The second proposed method is a kernel-based scheme. The gray-level difference between the template and the inspection image in a small neighborhood window can be efficiently calculated from a kernel function. The weighted kernel value can then be used for discriminating defect points from the background. The kernel method is also fast enough when compared to the simple image subtraction method and very responsive to detect low-contrast defects. The third method is called hybrid method which combines the SPC with multiple templates and the kernel-based method. The hybrid method takes a number of defect-free images but instead of calculating the “mean” template, it uses these images to calculate the values of the kernel weighting function. The hybrid method is computationally fast and very responsive to the detection of low-contrast defects.
All these proposed methods are compared against a benchmark method which is based on image subtraction for defect detection. The benchmark method calculates the minimum value of the pixel intensities between a template image and an inspection image. The minimum value is calculated in a neighborhood window and is used to overcome minor misalignments.
Tests are conducted with defect-free and defective PCBs and LED chips under different misalignments and illumination variation. In comparison with the benchmark method, all the proposed methods produce noise-free inspections results. They are computationally fast for on-line real-time defect detection and robust enough to overcome noise due to misalignments, changes in illumination and manufacturing product variation.
Automated inspection of assembled printed circuit boards (PCB) is a requirement to ensure the quality of the product and to decrease manufacturing cost. An assembled PCB comprises complex conductive paths with different electronic components such as integrated circuits (IC), resistors and capacitors, which make template matching the only possible approach for automated optical inspection. Current template-based defect detection techniques use simple image subtraction to identify defects on assembled PCBs which relies on accurate image alignment of the reference and inspection images. If this cannot be achieved, noise points are generated around the edges. Changes in illumination are another factor to consider since image difference will yield false detections.
This study is aimed to develop three template-based approaches for defect detection and, especially, focuses on PCBs and light emitting diode (LED) chips. These methods are robust under small misalignments, changes in illumination and manufacturing variation. The first proposed method, named statistical process control (SPC) with multiple templates, uses statistical process control techniques. It takes a number of defect-free images to form a “mean” template, described by the gray-level mean and standard deviation of each individual pixel in the image. This method is faster than the simple image subtraction method, and is tolerant to noise due to misalignments, illumination changes and manufacturing product variation. The second proposed method is a kernel-based scheme. The gray-level difference between the template and the inspection image in a small neighborhood window can be efficiently calculated from a kernel function. The weighted kernel value can then be used for discriminating defect points from the background. The kernel method is also fast enough when compared to the simple image subtraction method and very responsive to detect low-contrast defects. The third method is called hybrid method which combines the SPC with multiple templates and the kernel-based method. The hybrid method takes a number of defect-free images but instead of calculating the “mean” template, it uses these images to calculate the values of the kernel weighting function. The hybrid method is computationally fast and very responsive to the detection of low-contrast defects.
All these proposed methods are compared against a benchmark method which is based on image subtraction for defect detection. The benchmark method calculates the minimum value of the pixel intensities between a template image and an inspection image. The minimum value is calculated in a neighborhood window and is used to overcome minor misalignments.
Tests are conducted with defect-free and defective PCBs and LED chips under different misalignments and illumination variation. In comparison with the benchmark method, all the proposed methods produce noise-free inspections results. They are computationally fast for on-line real-time defect detection and robust enough to overcome noise due to misalignments, changes in illumination and manufacturing product variation.
Table of Contents

Abstract I
Table of Contents III
List of Figures V
List of Tables XII
1.Introduction 1
1.1 Problem Statement 1
1.2 Main Contributions 1
1.3 Objectives 3
1.4 Organization of the Thesis 3
2 Literature Review 5
2.1 Template Matching Methods 5
2.1.1 Cross-correlation 5
2.1.2 Normalized cross-correlation 6
2.1.3 Zero-Mean normalized cross-correlation 6
2.2 Current Defect Inspection Methods for PCB and IC Wafers 7
2.2.1 Simple image subtraction for defect detection 7
2.2.2 Defect detection using anisotropic kernels 7
2.2.3 Golden template 8
2.2.4 Meta-heuristics methods for PCB inspection 8
2.3 Summary 9
3 Defect Detection Using Image Subtraction 10
3.1 Introduction 10
3.2 Image Preprocessing of Inspection Images 11
3.2.1 Normalized cross-correlation 11
3.2.2 Transformation 13
3.2.3 Illumination normalization 14
3.3 Benchmark Method 17
3.3.1 Advantages and disadvantages 23
3.4 Statistical Process Control with multiple Templates 23
3.4.1 Advantages and disadvantages 29
3.5 Kernel-Based Defect Detection 29
3.5.1 Advantages and disadvantages 38
3.6 Hybrid Method 38
3.6.1 Advantages and disadvantages 40
3.7 Qualitative Comparison of the Proposed Defect Detection Methods 43
4 Experimental Results 45
4.1 System Configuration 45
4.2 Fiducial Setup for Alignment 46
4.3 Effects on Changes in Parameter Values 49
4.3.1 Statistical process control with multiple templates 49
4.3.1.1 Effect on the selection of the number of defect-free image samples 50
4.3.2 Kernel-based defect detection 56
4.3.2.1 Effect on the selection of the bandwidth 56
4.3.2.2 Effect on the selection of the neighborhood window 57
4.3.3 Hybrid Method 62
4.3.3.1 Effect on the selection of the bandwidth 62
4.3.3.2 Effect on the selection of the number of defect-free image samples 62
4.3.4 Evaluation of the defect detection methods for low-contrast defects 68
4.3.5 Measurements of efficacy for LED-chips inspection 70
4.4 Detection Results of the Benchmark and the Proposed Methods 75
4.4.1 Printed circuit board 1 75
4.4.2 Printed circuit board 2 76
4.4.3 Printed circuit board 3 76
4.4.4 LED wafers 77
4.5 Comparison of the Methods 83
5 Conclusions and Future Research 86
References 87
Appendix 89
A: Image Difference with Edge Verification 89
B: Defect Detection Results Using Different Noise Removal Window Sizes 101
C: LED-chips Defect Detection Results 125

List of Figures

Figure 3.1 Alignment and illumination normalization results on PCB and LED-chip defect-free samples. The template images (a1)-(e1) have an illumination value of 100 lux. The inspection images (a2) and (b2) have a value of 110 lux; images (c2) and (d2) have a value of 90 lux. Note in the template images the white squares represent the fiducial marks for template matching. 16

Figure 3.2 Evaluation of the simple image subtraction for defect detection with illumination variation. 17

Figure 3.3 Results from the benchmark method for defect-free PCB and LED-chip samples. Parameters used are a neighborhood window of 3×3 and a control limit constant K = 3. 21

Figure 3.4 Results from the benchmark method for defective PCB samples and LED-chip samples. Parameters used are neighborhood window of 3×3 and control limit constant K = 3. The defects in the images are: (a2) missing lead soldering; (b2) missing capacitor; (c2) PCB under-etching; (d2) PCB under-etching; (e2) broken LED chip. 22

Figure 3.5 SPC with multiple template detection results using 25 images as templates and a control constant K = 3. The benchmark method detection results has a neighborhood window of 3×3 pixels and a control limit constant of K = 3. The image in (a2) is a defect-free PCB; (b2) is a defective PCB with a missing lead soldering. 24

Figure 3.6 Results of statistics process control with multiple templates method in defect-free PCB and LED-chip samples. The number of defect-free images used in this test is 25 and a constant control limit K = 3. 27

Figure 3.7 Results of statistics process control with multiple templates method in defective PCB and LED-chip samples. The number of defect-free images used in this test is 25 and a constant control limit K = 3. The defects in the images are: (a2) missing lead soldering; (b2) missing capacitor; (c2) PCB under-etching; (d2) PCB under-etching; (e2) broken LED chip. 28

Figure 3.8 Graph of the kernel function under different exponent k values. 32

Figure 3.9 Kernel defect detection results using a bandwidth h* = 15. These detection results correspond to defect-free PCBs and LED-chip samples. Benchmark method uses a 3×3 neighborhood window and constant control limit K = 3. Kernel-based method uses a 3×3 neighborhood window, a bandwidth h* = 15 and constant control limit K = 3. 36

Figure 3.10 Kernel defect detection results using a bandwidth value equal to 15. These correspond to defective PCBs and LED-chip. Benchmark method uses a 3×3 neighborhood window and constant control limit K = 3. Kernel-based method uses a 3×3 neighborhood window, a bandwidth h* = 15 and constant control limit K = 3. The defects in the images are: (a2) missing lead soldering; (b2) missing capacitor; (c2) PCB under-etching; (d2) PCB under-etching; (e2) broken LED chip. 37

Figure 3.11 Evaluation of the hybrid method for defect-free images with misalignments and illumination changes. Bandwidth value is manually selected. Benchmark method uses a 3×3 neighborhood window and constant control limit K = 3. Hybrid method uses a bandwidth value of h = 9, 20 templates and constant control limit K = 3. 41

Figure 3.12 Evaluation of the hybrid method for defective images with misalignments and illumination changes. Bandwidth value manually selected. Benchmark method uses a 3×3 neighborhood window and constant control limit K = 3. Hybrid method uses a bandwidth value of h = 9, 20 templates and constant control limit K = 3. The defects in the images are: (a2) missing lead soldering; (b2) missing capacitor; (c2) PCB under-etching; (d2) PCB under-etching; (e2) broken LED-chip. 42

Figure 4.1A Different manufacturing defects on PCBs: (a) Missing component; (b) over- soldering bridged lead contacts on IC chip; (c) missing lead contacts on IC chip and (d) etching problems. 46

Figure 4.1B Different manufacturing defects on LED-chips: (a) Broken chip; (b) under-etching defect and (c) burn marks. 46

Figure 4.2 Example of a fiducial selection on the template image. The fiducials are represented in the solid lines and have a size of 50×50 pixel window. The dashed line represents the search limit for the fiducial in the corresponding inspection image. 47

Figure 4.3 LED wafer board: the image size of the wafer is 1260×960 pixels. Individual LED-chips have a size of 125×112 pixels. There are 48 individual LED-chips in this wafer. A LED-chip defect can be seen in the center of the wafer. 48

Figure 4.4 Product variations of LED-chips. (a) Irregular right edge, (b) irregular left edge, (c) irregular upper and left edges and (d) uneven right edge. 48

Figure 4.5 Example of selection of fiducial markers. The fiducials have a size of 25×25 pixels window. The dashed line represents the limits for the search of the fiducials. The search limit has a window size of 50×50 from the center of the fiducial. 49

Figure 4.6 Evaluation of the number of defect-free training samples in SPC with multiple templates for defect-free PCBs with 3×3 noise removal window and constant control limit K = 3 without illumination normalization. 52


Figure 4.7 Evaluation of the number of defect-free training samples in SPC with multiple templates for defect-free PCBs with 3×3 noise removal window and constant control limit K = 3 with illumination normalization. 53

Figure 4.8 Evaluation of the number of defect-free training samples in SPC with multiple templates for defective PCBs with 3×3 noise removal window and constant control limit K = 3 without illumination normalization. 54

Figure 4.9 Evaluation of the number of defect-free training samples in SPC with multiple templates for defective PCBs with 3×3 noise removal window and constant control limit K = 3 with illumination normalization. 55

Figure 4.10 Evaluation of different bandwidth values in the kernel-based method for defect-free PCBs with 3×3 neighborhood window, 3×3 noise removal window size and constant control limit K = 3. 58

Figure 4.11 Evaluation of different bandwidth values in the kernel-based defect detection method for defective PCBs with 3×3 neighborhood window, 3×3 noise removal window size and constant control limit K = 3. 59

Figure 4.12 Evaluation of different neighborhood window sizes in the kernel-based defect detection method for defect-free PCBs using h*=15, 3×3 noise removal window size and constant control limit K = 3. 60

Figure 4.13 Evaluation of different neighborhood window sizes in the kernel-based defect detection method for defective PCBs using h*=15, 3×3 noise removal window size and constant control limit K = 3. 61

Figure 4.14 Evaluation of the number of 20 defect-free training samples in the hybrid method for defect-free PCBs with different bandwidth values. (3×3 noise removal window size and constant control limit K = 3). 64

Figure 4.15 Evaluation of the number of 20 defect-free training samples in the hybrid method for defective PCBs with different bandwidth values. (3×3 noise removal window size and constant control limit K = 3). 65

Figure 4.16 Evaluation of the number of defect-free training samples in the hybrid method for defect-free PCBs with a bandwidth h = 6. (3×3 noise removal window size and constant control limit K = 3). 66

Figure 4.17 Evaluation of the number of defect training samples in the hybrid method for defective PCBs with a bandwidth h = 6. (3×3 noise removal window size and constant control limit K = 3). 67

Figure 4.18 Evaluation of the proposed defect detection methods for low-contrast defects using the recommended parameters. The difference of gray-level values between the adjacent background region of the defect and the defective region for (a2) is 6; (b2) is 7; (c2) is 5; (d2) is 10 and for (e2) is 12. A 3×3 noise removal window size and constant control limit K = 3 are used for the detection. 72

Figure 4.19 Evaluation of the proposed defect detection methods for low-contrast defects using the recommended parameters. The difference of gray-level values between the adjacent background region of the defect and the defective region for (a2) is 6; (b2) is 7; (c2) is 5; (d2) is 10 and for (e2) is 12. A 3×3 noise removal window size and constant control limit K = 2 are used for the defect. 73

Figure 4.20 Evaluation of the kernel-based method for low-contrast defects using a smaller bandwidth value. A manual selection is made on the bandwidth parameter h with a 3×3 and 5×5 neighborhood window respectively. A control constant of K = 3 is used for the detection. 74

Figure 4.21 Evaluation of the hybrid method for low-contrast defects using different bandwidth values and a 3×3 noise removal window. A control constant of K = 3 is used for the detection. 75

Figure 4.22 Test images for PCB1 under benchmark and proposed defect detection methods. All methods use a 3×3 noise removal window and a control limit of K = 3. 78

Figure 4.23 Test images for PCB2 under benchmark and proposed defect detection methods. All methods use a 3×3 noise removal window and a control limit of K = 3. 79

Figure 4.24 Test images for PCB3 under benchmark and proposed defect detection methods. All methods use a 3×3 noise removal window and a control limit of K = 3. 80

Figure 4.25 Test images for defect-free LED wafers under the benchmark and the proposed defect detection methods. All methods use a 3×3 noise removal window and a control limit of K = 3. 81

Figure 4.26 Test images for defective LED chips under the benchmark and the proposed defect detection methods. All methods use a 3×3 noise removal window and a control limit of K = 3. 82

Figure A.1 Canny Edge detection results extracted from template images. Images (a2)-(c2) show the canny edge detection process. (a3)-(c3) shows the detection results of the image subtraction method without the removal of the edges. Images (a5)-(c5) show that with the edge removal process, noise points due to misalignments can be overcome. A 3×3 dilation and noise removal window size and control limit K = 3 are used for the detection. 90

Figure A.2 Results of image subtraction with edge verification with defect-free PCB. Noise points represented in (a3), (b3), (c3) and (d3) corresponds not to the edges, this due to illumination change in certain areas. A 3×3 dilation and noise removal window size and control limit K = 3 are used for the detection. 94

Figure A.3 Results of image difference with edge verification with defective PCB and LED-chip samples. A 3×3 dilation and noise removal window size and control limit K = 3 are used for the detection. 95

Figure A.4 Image difference with edge registration inspection on defect-free samples without illumination normalization under different dilation window sizes. A 3×3 noise removal window size and control limit K = 3 are used for the detection. 97

Figure A.5 Image difference with edge registration inspection on defect-free samples with illumination normalization under different dilation window size. A 3×3 noise removal window size and control limit K = 3 are used for the detection. 98

Figure A.6 Image difference with edge registration inspection on defective samples without illumination normalization under different dilation window size. A 3×3 noise removal window size and control limit K = 3 are used for the detection. 99

Figure A.7 Image difference with edge registration inspection on defective samples with illumination normalization under different dilation window size. A 3×3 noise removal window size and control limit K = 3 are used for the detection. 100

Figure B.1 Benchmark method on defect-free samples. Results are shown using 3×3, 5×5 and 7×7 window neighborhoods without illumination normalization. A constant control K = 3 is used for the detection. 101

Figure B.2 Benchmark method on defective samples. Results are shown using 3×3, 5×5 and 7×7 window neighborhoods without illumination normalization. A constant control K = 3 is used for the detection. 102

Figure B.3 Benchmark method on defect-free samples. Results are shown using 3×3, 5×5 and 7×7 window neighborhoods with illumination normalization. A constant control K = 3 is used for the detection. 103

Figure B.4 Benchmark method on defective samples. Results are shown using 3×3, 5×5 and 7×7 window neighborhoods with illumination normalization. A constant control K = 3 is used for the detection. 104

Figure B.5 SPC with multiple templates defect-free detection results with 5×5 noise reduction window without illumination normalization. Column containing the template images just shows an example of one of the image used in the set of training images. A number of 15 training samples used in these results. 105


Figure B.6 SPC with multiple templates defect-free detection results with 7×7 noise removal window without illumination normalization. 106

Figure B.7 SPC with multiple templates defect-free detection results. Column containing the template images just shows an example of one of the image used in the set of training images. A number of 15 training samples used in these results. Illumination-normalized inspection images are used in this test. A 5×5 noise removal window is used. 107

Figure B.8 SPC with multiple templates defect-free detection results using a 7×7 noise removal window with illumination normalization. A 5×5 noise removal window is used for the detection. 108

Figure B.9 SPC with multiple templates defect-free detection results. Column containing the template images just shows an example of one of the image used in the set of training images. A number of 20 training samples used in these results. A 5×5 noise removal window is used without illumination normalization. 109

Figure B.10 SPC with multiple templates defect-free detection results using a 7×7 noise removal window without illumination normalization. A number of 20 training samples used in these results. 110

Figure B.11 SPC with multiple templates defect-free detection results using a 5×5 noise removal window with illumination normalization. A number of 20 training samples used in these results. 111

Figure B.12 SPC with multiple templates defect-free detection results using a 7×7 noise removal window with illumination normalization. A number of 20 training samples used in these results. 112

Figure B.13 SPC with multiple templates defect-free detection results. Column containing the template images just shows an example of one of the image used in the set of training images. A number of 25 training samples used in these results. A 5×5 noise removal window is used in the results without illumination normalization. 113

Figure B.14 SPC with multiple templates defect-free detection results using a number of 25 defect-free samples and 7×7 noise removal window without illumination normalization. 114

Figure B.15 SPC with multiple templates defect-free detection results using a number of 25 defect-free image samples and 5×5 noise removal window is used in the results with illumination normalization. 115

Figure B.16 SPC with multiple templates defect-free detection results using 25 defect-free samples and 7×7 noise removal window with illumination normalization. 116

Figure B.17 Kernel-Based defect detection results for defect-free images without illumination normalization. Using a bandwidth h* = 15, 3×3 neighborhood window and control limit K = 3. 117

Figure B.18 Kernel-Based defect detection results for defective images without using illumination normalization. Using a bandwidth h* = 15, 3×3 neighborhood window and control limit K = 3. 118

Figure B.19 Kernel-Based defect detection results for defect-free images. Using a bandwidth h* = 15, 3×3 neighborhood window and control limit K = 3. 119


Figure B.20 Kernel-Based defect detection results for defective images using illumination normalization. Using a bandwidth h* = 15, 3×3 neighborhood window and control limit K = 3. 120

Figure B.21 Hybrid method defect detection results on defect-free samples. Using a bandwidth h = 6, 20 images and control limit K = 3 without illumination normalization. 121

Figure B.22 Hybrid method defect detection results on defective samples. Using a bandwidth h = 6, 20 images and control limit K = 3 without illumination normalization. 122

Figure B.23 Hybrid method defect detection results on defect-free samples. Using a bandwidth h = 6, 20 images and control limit K = 3 with illumination normalization. 123

Figure B.24 Hybrid method defect detection results defective samples. Using a bandwidth h = 6, 20 images and control limit K = 3 with illumination normalization. 124

Figure C.1 LED-chip detection results on defect-free images. 125

Figure C.2 LED-chip detection results on defective images. 126

List of Tables

Table 3.1 Summary of the benchmark and proposed methods: advantages and disadvantages from a qualitative point of analysis. Bold entries are the best approaches for that property. 44

Table 4.1 System configuration for the personal computer used for this research. 45

Table 4.2 Computation time for the PCB alignment process. 48

Table 4.3 Computation time for the LED-chip alignment process. 49

Table 4.4 Computation time for the SPC with multiple templates method. 56

Table 4.5 Computation time of the kernel-based defect detection method. 57

Table 4.6 Computation time of the hybrid method. 58

Table 4.7 Summary of the effectiveness of the benchmark and the proposed methods on LED chips for defect detection. 59
Brown, L.G., 1992, A survey of image registration techniques, ACM Computing Surveys, vol. 24, pp. 325–376 .

Brunelli, R., Poggio, T., 1993, Face recognition: features versus templates, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 15, pp. 1042-1052.

Barnea, D.I., Silverman, H.F., 1972, A class of algorithms for fast digital image registration, IEEE Transactions Computers, vol. 21, pp. 179-186.

Chin, R.T., Harlow, C.A., 1982, Automated visual inspection: a survey, IEEE Transactions on Pattern Intelligence , vol. 4, pp. 557-573.

Di Stefano, L., Mattoccia, S., Tombari, F., 2005, ZNCC-based template matching using bounded partial correlation, Pattern Recognition Letters, vol. 26, pp. 2129-2134.

Duda, R. O., Hart, P. E., 1973, Pattern Classification and Scene Analysis, John Wiley and Sons, New York, NY.

Giachetti, A., 2000, Matching techniques to compute image motion, Image and Vision Computing, vol.18, pp. 247-260.

Krattenthaler, W., Mayer, K.J., Zeiler, M., 1994, Point correlation: A reduced-cost template matching technique, 1st IEEE Int. Conf. on Image Processing, Austin, Texas, USA, pp. 208–212.

Krispin, A.J. Rankov, V., 2006, Automated inspection of PCB components using a genetic algorithm template-matching approach, International Journal of Advanced Manufacturing Technology, vol. 35, pp. 293-300.

Lee, D.T., 1978, A computerized automatic inspection system for complex printed thick film patterns, Proceedings of Application of Electronic Imaging Systems, vol. 143, pp. 172-177.

Lewis, J.P., Fast normalized cross-correlation. Available online:
.

Rosenfeld, A., Vanderburg, G.J., 1977, Coarse-fine template matching, IEEE Transactions Systems Manufacturing Cybernetics, vol. 7, 104–197.

Shankar, N., Zhong, Z., 2005, Defect detection on semiconductor wafer surfaces. Microelectronic Engineering, vol. 77, 337–346.

Sun, C., 2002, Fast optical flow using 3D shortest path techniques, Image Vision Computing, vol. 20, pp. 981–991.

Tsai, D.M., Lin, C.T., 2003, fast normalized cross-correlation for defect detection, Pattern Recognition Letters, vol. 24, pp. 2625–2631.

Wu, W.Y. Wang, J. Liu, C.M.,1996, Automated inspection of printed circuit boards through machine vision, Computers in Industry, vol. 28, pp 103-111.

Xie, P. Guan, S.U., 2000, A golden template self-generating method for patterned wafer inspection, Machine Vision and Applications, vol. 12, pp. 149-156.

Zhao, F. Huang, Q. Gao, W., 2006, Image matching by normalized cross-correlation, IEEE International Conference on Acoustics, Speech and Signal Processing, vol. 2, pp. 729-732.

Zontak, M., Cohen, I., 2008, Defect Detection Using Anisotropic Kernels, Machine Vision and Applications, vol. 9, pp. 137-144.
QRCODE
 
 
 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                               
第一頁 上一頁 下一頁 最後一頁 top