After decades of research, it is exciting to see that face recognition technology has entered a most flourishing era. Driven by the latest development in data science and especially technical evolutions in computer vision and pattern recognition, face recognition has achieved significant progress over the last three years. In the near future, people can expect many useful and interesting face recognition applications to be deployed in many situations: they can be used for identifying suspects, organizing your photos with family and friends, and making computers better understand human beings. Many mysterious face recognition tricks depicted in movies may become reality in several years' time.
This thesis focuses on the development of face recognition algorithms that identify people from a single still image. Two questions are specifically studied. First, it introduces how we identify faces captured in controlled scenarios with cooperative users. In this scenario, a face recognition system captures a face and finds the most similar face from the ones stored in the face recognition system. Second, it describes our solutions for predicting face attributes from faces captured under arbitrary imaging conditions. These two problems were tackled by different schools of technologies: the solution to the first question employed a learning-free approach, whereas the latter question was solved by using the most recent Deep Learning technology. Thus, this thesis also reflects the technological evolution of face recognition over recent years.
To identify faces in controlled scenarios, we propose a novel Block Matching approach, which can effectively match faces without feature engineering or any machine learning components. By representing faces with very concise Gabor phase codes and matching them through our Block Matching approach, the identification accuracy is entirely comparable to and even better than the state-of-the-art. For predicting the attributes from faces captured in the wild, we propose leveraging the off-the-shelf mid-level representations from pre-trained convolutional neural networks. Comparative experiments show that our solution outperforms the previous state-of-the-art solution with a large margin in terms of both accuracy and efficiency.
The approaches described in this thesis may look different from the ``mainstream''. But, together with the empirical findings, I hope they could provide some insights and update widely adopted concepts for solving related face recognition and computer vision problems.
Stockholm: KTH Royal Institute of Technology, 2016. , 27 p.
Li, Haibo, Prof.Sullivan, Josephine, Prof.