臺灣博碩士論文加值系統

Recent technological advances in manufacturing industry allow high-speed mass production and manufacturers often utilize the online inspection machines to monitor the production process. However, the online inspection machines for image data in a high-speed mass production process are still in the development stage. They might cause misclassifications of the non-defective or defective products, and therefore, result in unnecessary process adjustment or profit loss. Previous studies developed some high accurate discriminant methods such as feature extraction in relation to a filtering window, and ensemble classifiers for identifying the defective products. However, these methods have some drawbacks which may cause high misclassification rates when they are employed in a high-speed mass production process. Hence, the objective of this study is to build a more precise defect detection system to discriminate the non-defective products and defective products using the convolutional neural network and experimental designs. The proposed method is verified to be superior to the existing defect detection system for classifying the image data in a high-speed mass production process.

Abstract i
Contents ii
List of tables iv
List of figures v

Chapter 1. Introduction 1
1.1 Research background 1
1.2 Research motivation 1
1.3 Thesis organization 3

Chapter 2. Literature review 4
2.1 Target product 4
2.2 Artificial defective products 5
2.3 Classification with features extracted from image data 5
2.4 The four drawbacks on conventional process 5
2.4.1 The drawback about converting polar coordinates 6
2.4.2 The drawback about preprocessing and feature selection 6
2.4.3 The drawback about classification 7
2.4.4 The drawback about characteristic of the inspection machine’s camera 7

Chapter 3. Research method 9
3.1 Data augmentation 9
3.2 Preprocessing 10
3.2.1 Filtering window 10
3.2.2 Gaussian filter 11
3.2.3 Laplacian filter 12
3.2.4 Sharpening filter 12
3.3 Convolutional neural network 12
3.3.1 Convolutional layer 15
3.3.2 Max-pooling layer 16
3.3.3 Classification layer 17
3.3.4 Dropout 17

Chapter 4. Experiments 19
4.1 Optimization of preprocessing methods with experimental designs 19
4.2 Deep residual learning 20
4.3 Experiments with L8 orthogonal array 23
4.4 Experiment results 25

Chapter 5. Discussion 31
5.1 Consideration of the optimal combination 31
5.2 Comparison of the accuracy with conventional procedure 32
5.3 When to stop training the ResNet 34

Chapter 6. Conclusion 37

References 39

1. Y. LeCun et al., Backpropagation applied to handwritten zip code recognition, Neural Computation, Vol.1, pp.541-551, 1989.
2. K. Fukushima, Neocognitron. A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position, Vol.202, 1980.
3. Y. LeCun et al., Gradient-based learning applied to document recognition, Proc.IEEE, Vol.86, No.11, pp.2278-2324, 1998.
4. D. Cireşan et al., Multi-Column Deep Neural Network for Traffic Sign Classification, Neural Networks, Vol.32, pp.333-338, 2012.
5. G. E. Hinton et al., Improving neural networks by preventing co-adaptation of feature detectors, CoRR, vol.abs/1207.0580., 2012.
6. N. Srivastava et al., Dropout: A simple way to prevent neural network from overfitting. Journal of Machine Learning Research, 15:1929-1958, 2014.
7. I. J. Goodfellow et al., Maxout Networks, Proceeding on the 30th International Conference on Machine Learning, JMLR: W&CP volume28, 2013.
8. V. Nair et al., Rectified linear units improve restricted Boltzmann machines, Proceedings of ICML, 2010.
9. Rahaman et al., Automatic defect detection and classification technique from image: a special case using ceramic tiles, International Journal of Computer Science and Information and Security, 1(1):22-33, 2009.
10. M.Arun et al., Automatic defect detection of steel products using supervised classifier, International Journal of Innovative Research in Computer and Communication Engineering, 2(3):3630-3635, 2014.
11. A. Kumar, Neural network based detection of local textile defects, Pattern Recognition, 36(7):1645-1659, 2003.
12. K. Yoshida et al., Defect Detection for image Data using Image processing and Machine Learning, Proceeding of Asian Network for Quality Congress 2016, 2016.

13. N. Sano et al., Defect Detection using two-dimensional moving range filter and unanimous vote among color component classifiers, Proceedings of the 3rd multidisciplinary International Social Networks Conference, 2015.
14. T. Iwasawa et al., Defect Detection from image data with feature extraction using orthogonal array, Proceeding of Asian Network for Quality Congress 2016, 2016.
15. X. Glorot et al., Deep Sparse Rectifier Neural Networks, Proceedings of AISTAT, 2011.
16. D. Maar et al., Theory of Edge Detection, Proceeding of the Royal Society of London. Series B, Biological Sciences, Vol. 207, No. 1167, pp. 187-217, 1980.
17. A. Huertas et al., Detection of Intensity Changes with Subpixel Accuracy Using Laplacian-Gaussian Masks, “IEEE Trans. On Pattern Analysis and Machine Intelligence, Vol. 8, No. 5, pp. 651-664, 1986.
18. M. Takagi et al., Handbook of Image Analysis, Tokyo: University of Tokyo Press, 2004, (in Japanese).
19. K. He et al., Deep Residual Learning for Image Recognition, In CVPR, 2015.
20. K. Simonyan et al., Very Deep convolutional networks for large-scale image recognition, In ICLR, 2015.
21. T. Okatani, Deep Learning, Kodansya press, 2015, (in Japanese).
22. A. Canziani et al., An Analysis of Deep Neural Network Models for Practical Applications, arXiv present arXiv:1605.07678v4, 2017.
23. C. Szegedy et al., Inception-v4, inception-resnet and the impact of residual connections on learning, arXiv present arXiv:1602.07261, 2016.