PUKYONG

GUIDE OAK 국가리포지터리 부경대학교

검색

PUKYONG Repository 대학원 제어계측공학과

Metadata Downloads

Alternative Title: Design of Activation Function in Deep Learning Model for Image Object Detection

Abstract: Trams are considered as next-generation public transportation driven by electricity as eco-friendly energy. The subway is also operated by electricity, but the tram has the advantage of being easier to install tracks than the subway. Also, it is installed on the ground, so anyone can easily access it. In this paper, the experimental purpose is finding out a more suitable dataset for object recognition in trams, and what performance changes the self-developed Rish activation function makes in the object recognition model. To identify, the experiment was conducted as follows. First, each YOLOv4 model was trained with different datasets with different classes, respectively. The YOLOv4 models were compared to which dataset is easier to learn and compared to which dataset is more suitable for developing an object recognition model for autonomous driving through a camera installed in a tram. The Rish activation function and the Mish activation function used by the original YOLOv4 model were compared in the same way as the dataset experiment.

URI: https://repository.pknu.ac.kr:8443/handle/2021.oak/32807
http://pknu.dcollection.net/common/orgView/200000643368

Table Of Contents: 1. 서 론 1
1.1 연구배경 및 필요성 1
1.2 논문의 구성 4
2. 인공신경망 이론 및 객체 인식 모델 기법 5
2.1 객체 인식 모델 5
2.2 YOLOv4 9
2.2.1 데이터 증강 기법 9
2.2.2 아키텍처 13
2.2.3 손실 계산 21
2.2.4 최적화 함수 26
2.2.5 과적합 방지 및 정규화 기법 28
2.2.6 그 외 기법 32
2.3 활성화 함수와 Rish 36
2.3.1 Sigmoid와 쌍곡선탄젠트 36
2.3.2 ReLU와 ReLU 응용 활성화 함수 38
2.3.3 Rish 42
2.4 데이터셋 48
3. 실험 방법 52
3.1 딥러닝 모델 학습 방법 52
3.2 딥러닝 모델 성능 지표 54
3.3 YOLOv4모델을 이용한 트램 환경에서의 객체 인식 실험 61
4. 실험 결과 66
4.1 데이터 셋 차이에 의한 YOLOv4 모델 학습 결과 66
4.2 YOLOv4모델에 Rish를 적용한 결과 71
5. 결과 분석 75
5.1 데이터 셋에 의한 YOLOv4모델의 정량적 성능 차이 분석 75
5.2 Rish을 적용한 딥러닝 모델의 정량적 성능 분석 77
5.3 영상을 통한 트램 환경에서의 객체 인식 성능 평가 81
6. 결론 89
참고문헌 91
감사의 글 96

Authorize & License

qrcode