近年來，利用機器學習或深度學習實現人工智能，已廣泛地應用在各領域，特別在影像資料的分類與辨識問題。訓練深度網路模型仰賴大規模訓練數據集，才能達到好的學習效果，並在執行新的分類任務時，能獲致準確的辨識判斷。 然而，許多領域無法取得大型數據集，例如醫學圖像通常需高昂的成本，大資料取得不易，另一個例子是工業 PCB 和焊接 AOI 圖像，在產品開發早期亦沒有足夠的圖像文件來訓練複雜的模型。此時，資料集擴增技術為小資料集及資料有限的情境下執行深度學習任務提供一個解決方案。經由適當的資料擴增技術，能增強特定影像特徵，提昇後續深度學習分類預測成效。
此論文中探究數種資料集擴增技術，涵蓋了影像的幾何對稱轉換和形態學轉換等，並實證應用於小數據集的醫學影像資料、PCB 圖像資料、AOI影像資料。本研究在convolution neural network 與 Resnet50 的模型架構下，進行了資料集擴增技術對這三類影像分類準確性的評估與比較。實驗結果顯示使用影像翻轉的技術(水平或垂直翻轉)，圖像分類的準確率在胸部 X-Ray 的數據集提昇了12% (由73.78% to 85.95%)，在PCB 表面數據集提昇15% (由29.03% 到 44.49%)，在AOI 數據集提昇了13% (由72.31% 到 85.11%)。利用形態學的資料轉換 (Open, Close, Hit-or-Miss, Gradient 運算) 將原始圖像轉換為新圖像以進行的資料擴增實驗，也都提升了這三類影像分類的準確性。尤以Close-operation的資料轉換在後續分類預測的改進效果最佳，對胸部 X 射線數據集的準確率從 73.78% 提高到 88.29%，對PCB數據集的準確率從 29.03% 提高到 37.14%，對AOI 數據集的準確率從 72.31% 提高到 90.31%。 總結以上，善用數據增強和特徵提取的圖像處理能有效輔助小數據集及不平衡分類在深度學習的實務應用。

AI has become a powerful tool in wide applications. In particular, machine learning and deep learning models have been very successful in a variety of image analysis tasks, ranging from classification problems, and segmentation identification to fault detections. To achieve good accuracy as like-human behaviors, the deep learning algorithm requires a large-scale training dataset. Such a large-scale dataset is not typically available in many fields of common casework. To overcome this issue, data augmentation is a popular low-cost approach to expanding data size and enhancing certain features in a small dataset scenario. Due to the nature of data types, different augmentation techniques can be designed and applied to emphasize unique data features in order to generate useful augmented data for specific target tasks.
This thesis studies the effectiveness of data augmentations to improve the image classifier accuracy of the small dataset. The data experiments are planned and implemented in the framework of CNN with ResNet50 network architecture for three datasets, including X-Ray medical image dataset, the PCB surface image dataset, and the Soldering AOI image dataset. Two types of data augmentation methods are considered. One is based on image flipping and the other is based on image morphology. Results show that the image flipping technique significantly improves the prediction accuracy for three datasets, ranging from 12% to 15%. The image morphology techniques (Open, Close, Hit-or-Miss, and Gradient operations) also improve the accuracy with the gain ranging from 5% to 18%. In particular, the close-operation has the greatest improvement among four morphological operations considered in this study. This study found that many traditional image processing techniques are very effective for data augmentation to enhance specific data features and extract data information, which benefit to deep learning tasks in small dataset and imbalanced data scenarios.

Contents
Chapter 1. Introduction----------------------------------1
1.1 Motivation and Contribution-----------------------1
1.2 Thesis Organization-------------------------------2
Chapter 2. Related Research------------------------------4
Chapter 3. Image Analysis With Neural Network------------6
3.1 CNN (Convolutional Neural Network) Architecture---6
3.1.1 Convolution layers---------------------------------7
3.1.2 Pooling layer:-------------------------------------7
3.1.3 Fully-connected layers:----------------------------8
3.1.4 Activation function--------------------------------9
3.2 ResNet--------------------------------------------9
3.2.1 ResNet Advantages---------------------------------10
3.2.2 ResNet Block Architecture-------------------------11
Chapter 4. Data Augmentation----------------------------12
4.1 Basic Image Manipulation-------------------------12
4.2 Image Processing Techniques Based on Morphology--13
4.3 Toolbox for Morphology Image Processing----------17
4.4 Proposed Workflow and Experiments----------------18
Chapter 5. Chest X-Ray Experiment Results---------------21
5.1 Chest X-Ray Images and Augmentations-------------21
5.2 Experiment Result--------------------------------24
5.3 Impact of Structure Element on Performance-------26
Chapter 6. PCB Image Experiment Result------------------28
6.1 PCB Surface Image Dataset------------------------28
6.2 PCB Dataset Image Experiment Result--------------31
6.3 Edge Transformed Result for PCB Surface Dataset--32
Chapter 7. Solder Image Experiment Result--------------33
7.1 AOI Solder Image Dataset-------------------------33
7.2 AOI Solder Dataset Experiment Result-------------34
Chapter 8. Conclusions---------------------------------37
8.1 Application Summary------------------------------37
8.2 Contributions and Future Research----------------39

References
[1] M. Agnieszka, G. Michał, "Data augmentation for improving deep learning in image classification problem", IEEE:International interdisciplinary PhD workshop, (2018), pp. 117-122.
[2] K. He, X. Zhang, S. Ren, J. Sun, "Deep residual learning for image recognition", IEEE:Conference on computer vision and pattern recognition, (2016), pp. 770-778.
[3] S. Ioffe, C. Szegedy, "Batch normalization: Accelerating deep network training by reducing internal covariate shift", PMLR:International conference on machine learning, volume 37, (2015), pp. 448-456.
[4] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A.A. Bharath, "Generative adversarial networks - An overview", IEEE:IEEE signal processing magazine, volume 35, (2018), pp. 53-65.
[5] X. Yi, E. Walia, P. Babyn, "Generative adversarial network in medical imaging: A review", arXiv: Medical image analysis, volume 58, (2019) pp. 3-9.
[6] D.I. Morís, J.J. de Moura Ramos, J.N. Buján, M.O. Hortas, "Data augmentation approaches using cycle-consistent adversarial networks for improving COVID-19 screening in portable chest X-ray images", Elsevier: Expert Systems with Applications, volume 185, (2021), pp. 2-6.
[7] W. Jin, S. Dong, C. Dong, X. Ye, "Hybrid ensemble model for differential diagnosis between COVID-19 and common viral pneumonia by chest X-ray radiograph", Elsevier: Computers in Biology and Medicine, volume 131, (2021), pp. 2-8.
[8] S. Dunford, S. Canumalla, P. Viswanadharn, "Intermetallic morphology and damage evolution under thermomechanical fatigue of lead (Pb)-free solder interconnections", IEEE: Electronic Componets and Technology Conference, (2004), pp. 1-11.
[9] W. Zhao, R. Chellappa, P.J. Phillips, A. Rosenfeld, "Face recognition: A literature survey", ACM: Computing surveys, volume 35, (2003), pp. 399-458.
[10] G. Wolberg, "Geometric transformation techniques for digital images: a survey", Columbia University: Computer Science, (1988), pp. 13-17.
[11] M. Elgendi, M.U. Nasir, Q. Tang, D. Smith, J.-P. Grenier, C. Batte, B. Spieler, W.D. Leslie, C. Menon, R.R. Fletcher, N. Howard, R. Ward, W. Parker, S. Nicolaou, "The Effectiveness of Image Augmentation in Deep Learning Networks for Detecting COVID-19: A Geometric Transformation Perspective", Frontiers: in Medicine, volume 8, (2021), pp. 1-12.
[12] R. Heriansyah, S.A.R.S.A. Bakar, M. Mun'im Ahmad Zabidi, "Segmentation of PCB Image Into Simple Generic Patterns Using Mathematical Morphology and Windowing Technique", Proceedings of National Conference: Computer Graphic & Multimedia, (2002), pp. 1-7.
[13] J. Naam, J. Harlan, S. Madenda, E.P. Wibowo, "Identification of the proximal caries of dental x-ray image with multiple morphology gradient method", ACADEMIA: International Journal on Advanced Science, Engineering and Information Technology, volume 6, (2016), pp. 343-346.
[14] K. O'Shea, R. Nash, "An introduction to convolutional neural networks", arXiv:1511.08458_Computer Science > Neural and Evolutionary Computing, (2015), pp. 1-11.
[15] D. Dai, "An Introduction of CNN : Models and Training on Neural Network Models", IEEE:International Conference on Big Data, Artificial Intelligence and Risk Management (ICBAR), (2021), pp. 1-4.
[16] S. Albawi, T.A. Mohammed, S. Al-Zawi, "Understanding of a convolutional neural network", IEEE: International conference on engineering and technology (ICET), (2017), pp. 1-6.
[17] A. Budhiman, S. Suyanto, A. Arifianto, "Melanoma Cancer Classification Using ResNet with Data Augmentation", IEEE: International Seminar on Research of Information Technology and Intelligent Systems, (2019), pp. 1-4.
[18] G. Bradski, "The openCV library: https://docs.opencv.org/3.4/d6/d00/tutorial_py_root.html", Dr. Dobb's Journal: Software Tools for the Professional Programmer, volume 25, (2000), pp. 120-123.
[19] C. Shorten, T.M. Khoshgoftaar, "A survey on image data augmentation for deep learning", Springer: Journal of big data, volume 6, (2019) pp. 12-22.
[20] R.C. Gonzalez, R. Woods, "Digital image processing 4th Edition", Pearson Education, (2018), pp. 635-652.
[21] G.C. Paul, C.R. Thomas, "Characterisation of mycelial morphology using image analysis", Springer: Advances in Biochemical Engineering, (1998), pp. 1-59.
[22] P. Bhattacharya, W. Zhu, K. Qian, "Shape recognition method using morphological hit-or-miss transform", SPIE Journal: Optical Engineering, volume 34, (1995), pp. 1-8.
[23] S. Gollapudi, "Learn computer vision using OpenCV: With Deep Learning CNNs and RNNs", Springer: Electronic, (2019), pp. 51-82.
[24] A.F. Villán, "Mastering OpenCV 4 with Python: a practical guide covering topics from image processing, augmented reality to deep learning with OpenCV 4 and Python 3.7", Packt Publishing Ltd, (2019), pp. 280-305.
[25] D. Krstinić, M. Braović, L. Šerić, D. Božić-Štulić, "Multi-label classifier performance evaluation with confusion matrix", University of Split: Computer Science & Information Technology, (2020), pp. 1-14.
[26] AIdea Collaboration Platform, "AOI Defect Classification", Open Source: "https://aidea-web.tw/topic/252eb73e-78d0-4024-8937-40ed20187fd8", (2021).