CNN 을 사용하여 단일 이미지에서 상위 수준의 헤어스타일 속성을 분류하고 회귀 학습
Learning to Classify and Regress High-Level Hair Attributes from a Single Image Using CNNs
- 주제(키워드) hair , deep convolutional neural network , deep learning , classification , regression , attributes , annotation , data generation
- 발행기관 아주대학교
- 지도교수 신현준
- 발행년도 2020
- 학위수여년월 2020. 2
- 학위명 석사
- 학과 및 전공 일반대학원 라이프미디어협동과정
- 실제URI http://www.dcollection.net/handler/ajou/000000029878
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
We present a fully automatic framework that classifies and regress the high-level hair attributes specifically length of hair and hair color from a single input image. This research was conducted to reduce the hard labor preparing the desired data for training and learning purpose. Also, this research can expand to different human face analysis research. We created new annotations for ten thousand image data. Which can be used for future research on human face analysis. We developed an automatic file distributor when annotations are liable. We further show the effectiveness of our experiment on cellular phone taken images.
more목차
I Introduction 1
II Related Works 4
A Segmentation Approach 4
B Detection Approach 5
C Convolutional Neural Networks 5
III Method 7
A Network Architectures 7
1 VGG-16 7
2 ResNet-20 8
B Pre-Processing 10
1 Classification Data-Generation 11
2 Regression Data-Generation 12
IV Experiment 19
A Loss Function 20
B Overall Training Process 22
V Result 24
A Hair Length Classification Prediction 24
B Hair Color Regression Prediction 27
VI Limitation and Discussion 33
VII Conclusion 34
References 35