Title Page
Contents
국문 초록 10
ABSTRACT 12
Ⅰ. Introduction 14
Ⅱ. The definition of diseases 16
1. Cataracts 16
2. Ulcerative keratitis 16
3. Pigmentary keratitis 17
4. Entropion of the eyelid 17
5. Eyelid Tumors 18
6. Epiphora 19
Ⅲ. Data Collection and Data Preprocessing 19
Ⅳ. Exploring Deep Learning Models 27
1. DenseNet 30
2. Xception 30
3. EfficientNet 32
4. Variational AutoEncoder 34
5. Generative Adversarial Networks 35
6. Transfer Learning 36
Ⅴ. Method and Results 38
1. Image Augmentation 38
2. Ensemble 39
3. One versus All 41
4. One versus One 45
5. Classification of latent vectors extracted using VAE 47
6. Image Anomaly Detection using GAN 49
Ⅵ. Conclusion 52
References 55
Table 1. Experimental Enviroment 27
Table 2. Performance metrics for each model 40
Table 3. Each binary classifier was evaluated using various performance metrics. 43
Figure 1. Early, immature, and mature cataracts 16
Figure 2. Ascending and descending ulcerative keratitis 17
Figure 3. Pigmentary keratitis 17
Figure 4. Entropion of the eyelid 18
Figure 5. Eyelid Tumors 18
Figure 6. Epiphora 19
Figure 7. Distribution of data by shooting device 20
Figure 8. Distribution of the Top 10 Dog Breeds 21
Figure 9. Number of Images per Disease 22
Figure 10. Overall image pre-processing Workflow 23
Figure 11. Hue, Saturation, Value Color Map 23
Figure 12. Image after division into H, S, and V regions 24
Figure 13. Image Histogram distribution 24
Figure 14. Applied Histogram Equalization 25
Figure 15. Example of Median Blur 25
Figure 16. Applied Median Blur 26
Figure 17. Convolution Layer 28
Figure 18. Pooling Layer 29
Figure 19. Fully Connected Layer 29
Figure 20. Dense Connection 30
Figure 21. Xception Architecture 31
Figure 22. The internal structure of each flow 31
Figure 23. Modified Depthwise Separable Convolution 32
Figure 24. The basic idea of EfficientNet (a) Example of baseline network (b) The compound scaling approach involves scaling the width, depth, and resolution of a... 33
Figure 25. VAE Architecture 35
Figure 26. GANs Architecture 36
Figure 27. The structure of the Fully Connected Layer 37
Figure 28. Image Augmentation Results 39
Figure 29. Example of Ensemble Model (Bagging) 39
Figure 30. Confusion Matrix of the Ensemble Model 41
Figure 31. Example of One vs All Method 42
Figure 32. Confusion Matrix of the classification using OVA 44
Figure 33. Example of One vs One 45
Figure 34. One versus One Confusion Matrix 46
Figure 35. Image classification process of VAE 48
Figure 36. Image Generated Using VAE 48
Figure 37. Image generated using GAN. 51