표제지
요약문
Abstract
목차
제1장 서론 12
1.1. 연구배경 12
1.2. 연구 내용 13
제2장 데이터 측정 시스템 개발 및 데이터 수집 15
2.1. 이미지 데이터 수집 시스템 제작 15
2.2. 이미지 데이터 수집 18
제3장 합성 곱 신경망을 이용한 분류 모델 제작 23
3.1. 합성 곱 신경망 기본 개요 23
3.2. 데이터 전처리 37
3.3. 분류 모델 학습 및 검증 과정 39
3.4. 모델 성능 개선 및 최적 모델 선정 41
제4장 도메인 적응 및 불량 상태 분류 53
4.1. 도메인 적응 개요 53
4.2. 데이터 증강을 통한 도메인 적응 기법 54
4.3. 불량 상태 분류 모델 61
제5장 연구 결론 67
제6장 참고문헌 69
Table. 2-1. Number of images per golf ball brand 19
Table. 2-2. Number of images, categorized by brand, featuring identifiable logos and those where logo identification is challenging 22
Table. 3-1. Number of images identifiable with class and images where class identification is challenging 37
Table. 3-2. Comparison Based on the Number of Channels in the Convolutional Layer 44
Table. 3-3. Comparison Based on Batch Normalization 46
Table. 3-4. Classification Accuracy per Class 50
Table. 4-1. Classification Accuracy per Class in the Target Domain 54
Table. 4-2. Number of augmented data in the domain based on λ 56
Table. 4-3. Classification Accuracy per Class in the target domain for the model adapted through domain adaptation. 59
Table. 4-4. Number of Images per Class for the Defect Classification Model 61
Table. 4-5. Number and types of data based on the dataset 62
Fig. 2-1. Image Data Collection System (a) Data collection system (b) Camera and camera mounting device... 17
Fig. 2-2. Examples of collected golf ball brands (a) Callaway (b) Srixon (c) Comet (d) Bridgestone (e) Kakao... 19
Fig. 2-3. Examples of balls with challenging brand recognition (a) Cases where the logo isn't visible (b) Cases where the logo is... 21
Fig. 2-4. Examples of image for the defective classification model (a) normal golf ball (b) a surface-contaminated golf bal... 21
Fig. 3-1. Neural Network Architecture 25
Fig. 3-2. Convolution Neural Network Architecture 26
Fig. 3-3. Example of Convolution Operation 27
Fig. 3-4. Example of Padding Process 29
Fig. 3-5. Example of Operation Process with Different Strides 30
Fig. 3-6. The process of computing feature maps for 2D data. 33
Fig. 3-7. The process of computing feature maps for 3D data. 34
Fig. 3-8. Pooling types and examples of their outcomes 35
Fig. 3-9. Fully Connected Layer Architecture 36
Fig. 3-10. Process for Determining Step-wise Learning Rates (a) First Step in Determining Learning Rates (b) Second Step in... 42
Fig. 3-11. Comparison of Validation Loss Based on the Number of Convolutional Layers 45
Fig. 3-12. Comparison of Validation Loss Based on the Number of Channels in the Convolutional Layer 45
Fig. 3-13. Comparison of Validation Loss Based on Batch Normalization 47
Fig. 3-14. Comparison of Validation Loss Based on Dropout Rates 47
Fig. 3-15. Training Loss and Validation Loss with the Selected Structure 49
Fig. 3-16. Validation Accuracy with the Selected Structure 49
Fig. 3-17. Confusion Matrix representing the evaluation results of the brand classification model on the test data 51
Fig. 3-18. Examples of misclassifications based on features other than the logo 52
Fig. 4-1. Confusion Matrix representing the evaluation results of the brand classification model on the target domain data. 55
Fig. 4-2. Training and validation losses during 10 epochs 58
Fig. 4-3. Validation accuracy during 10 epochs 58
Fig. 4-4. The confusion matrix representing the evaluation results of the domain adapted model on the target domain data 60
Fig. 4-5. Convolutional Neural Network Architecture for Defect Classification Model 63
Fig. 4-6. Training and validation losses during 15 epochs 65
Fig. 4-7. Validation accuracy during 15 epochs 65
Fig. 4-8. Confusion Matrix for Defect Classification Model by Class 66