Construction of Dataset for Face Identification in the Wild and Utilization of Public Face Datasets
- 주제(키워드) Face Dataset , Construction of Dataset , Face Identification , In-the-wild Environment
- 주제(DDC) 621.381
- 발행기관 아주대학교
- 지도교수 허용석
- 발행년도 2023
- 학위수여년월 2023. 8
- 학위명 박사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000032857
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
In recent years, deep learning advancements have contributed to the progress of various fields, including object detection and face recognition. Consequently, the domains of face detection and identification have also witnessed substantial advancements. However, researchers face significant challenges in acquiring suitable datasets for their research. In researching deep learning, a dataset is one of the most essential elements. A large amount of high-quality data is necessary for training and testing deep learning algorithms, and good datasets can significantly impact algorithm performance. Researchers usually have to collect their data or reprocess existing datasets to acquire appropriate data for their research. The significance of facial datasets and the necessity of employing suitable database construction methods to improve the accuracy of facial-related research are demonstrated in this thesis. Firstly, the thesis performs a classification and analysis of existing face datasets to explore the strengths, weaknesses, and characteristics of each dataset. Secondly, an Image Recommendation System is proposed, which generates customized images according to the user’s psychological state, thereby highlighting the utilization of existing datasets. Lastly, the research focuses on the construction of a dataset and the development of its construction methods for identifying the faces of preschool children in challenging, wild environments. In this thesis, a classification of publicly available face datasets is presented for researchers to utilize. While there are commonly used face datasets, researchers often resort to collecting their own data or processing existing datasets to obtain suitable data for their research purposes. The selection of an appropriate face database is crucial for researchers in the field of face recognition, computer vision, and related disciplines. Currently, the availability of face datasets has diminished due to the reinforcement of laws pertaining to the protection of personal information. Consequently, datasets containing personal details such as names, genders, and ages have become scarce. Furthermore, certain public datasets have been converted into private ones, posing a challenge for many researchers in accessing the datasets they require. Public datasets containing personal information are often limited to well-known individuals, and datasets that offer a broader range of information are typically available for a fee. While paid datasets generally exhibit good quality, they may be unaffordable for many researchers, thereby limiting their accessibility. In this thesis, the face datasets have been categorized based on factors such as their availability (public or distributed), collection environment, and provided annotation information. To build the recommender system, two public datasets have been utilized: the CK+ dataset for facial expression recognition and the MORPH database for age and gender estimation. For the classification tasks, the CNN (Convolutional Neural Network) algorithm has been employed as the classifier, and the VGG16 model architecture has been utilized. Environmental data has been collected using smartphone sensors, RSS (Really Simple Syndication) from the Korea Meteorological Administration, and data from OpenWeatherMap. In this thesis, CycleGAN has been used to trans ii form images of natural environments into biometrically appropriate images based on the estimated state of the subject. The system has performed real-time age, gender, and facial expression estimation and recognition based on the images captured by the camera. Facial expression recognition has achieved an accuracy of approximately 89% across eight expression classes. The overall accuracy of age estimation has been around 84%, and the overall accuracy of gender estimation has been approximately 98%. These facial expressions, age, and gender information have been combined with internal and external environmental data. Using the integrated data and the findings from color psychology and environmental psychology, the images are transformed into colors that best represent the user’s psychological state. In this thesis, the methodology for constructing a dataset targeting minors is described. Facial data of children aged 2 to 7 has been collected to build the dataset. A system has been developed using the collected dataset to recognize children’s faces in wild environments. However, to construct a facial recognition system in a real-world, in the-wild environment, a new dataset needs to be created. In this thesis, a facial recognition system has been developed specifically for children aged 2 to 7. The availability of children’s data has been severely limited, and its usage has been restricted. Legal protection for a child’s privacy has prioritized parental control and consent. The privacy rights of children should not have been conditioned upon the desires, actions, or control of others. Therefore, even with parental consent, all information has been kept confidential to safeguard the child’s privacy. Due to these legal and ethical concerns, there has been a scarcity of publicly available datasets for children’s facial recognition. Considerations and methods for collecting a database of children’s data have been presented, and effective camera installation positions for children’s facial recognition have been proposed. To achieve this, facial data has been collected from a total of 74 children aged 2 to 7 in actual daycare facilities, and experiments have been conducted by installing cameras in various environments. Through these experiments, the reliability of the collected data has been enhanced using the proposed methods. Based on the experimental results, the optimal position for the camera in the experimental space has been found to be approximately 90cm away from the door. At this location, the maximum distance for face recognition of children was 2m70cm, with an accuracy of 70.82%.
more목차
I Introduction 1
I.A. Purpose of use of face data 1
I.B. Limitations of Face Dataset 3
I.C. Select good dataset 4
I.D. Classification of the dataset 8
I.D.1. Classification of datasets based on public availability and consent of subjects 8
I.D.2. Classification of datasets based on the collection environment 10
I.D.3. Classifying the dataset based on the annotations 12
I.D.4. Dataset classification based on the dataset size 13
II Face dataset 16
II.A. Datasets that provide personally identifiable information 20
II.B. Datasets that provide personal information 27
II.C. Datasets that provide only images and brief information 36
III Image Recommendation System Using Existing Dataset 50
III.A. Recommendation system 51
III.B. Human psychology based on environmental factors 51
III.C. Facial expression recognition, Age & Gender estimation 62
III.D. Color psychology & Image colorization 68
III.E. Image recommendation system 73
III.F. Questionnaire and results 79
III.F.1. Pre-questionnaire 79
III.F.2. Last questionnaire 82
III.G. Differences from existing recommendation systems 89
III.H. Results of building a system using an existing dataset 92
IV Collecting and Recognizing Face Data in Preschool Children 94
IV.A. Dataset containing minors data 94
IV.B. Precautions for Collecting Child Databases 98
IV.C. Children's face data collection 102
IV.C.1. The preparatory process for data collection 103
IV.C.2. Database structure 104
IV.C.3. Training data collection method 106
IV.C.4. Choosing the location of daycare center cameras 110
IV.D. Algorithm 128
IV.D.1. Face Detection: MTCNN, RetinaFace 128
IV.D.2. Face Alignment & Face Landmark Estimation 133
IV.D.3. Face Identification: ArcFace, Nearest Neighbor 134
IV.E. Experiment 137
IV.E.1. Pre-Experiment for DB Performance Test 137
IV.E.2. Evaluation of face detection performance according to changes in video resolution 140
IV.E.3. Recognition rate and speed difference according to the embedding data value 141
IV.E.4. Child Face Recognition Rate 143
IV.F. Result 148
V Conclusion 151
