Face Recognition Using Feature Fusion and Deep Learning

  • Xin Wei

Student thesis: Doctoral Thesis

Abstract

Face recognition is a research hotspot in the fields of computer vision and deep learning. Over the past decade, impressive progress has been made, but there are still challenges in face recognition, as the facial portion of an image can easily be affected by various factors like occlusion, age, illumination and pose. A typical face recognition system includes face detection, face image representation and final classification, where the key process is face image representation, namely, face image-based feature extraction. This thesis focuses on learning highly discriminative facial features with traditional machine learning techniques and deep learning techniques to improve face recognition performance.
With traditional machine learning techniques, we firstly propose a feature fusion method called Multi-descriptor Fusion (MDF) in Chapter III to generate hyper-high dimensional descriptor features and highly discriminative features. The performance of MDF is competitive even compared with some deep learning-based methods. In a deep neural network, a loss function plays an extremely important role, which supervises the whole training process and offers feedback information to optimise the parameters of the neural network. Better loss functions can lead to more effective features. Therefore, we propose four loss functions in deep learning-based face recognition, where Minkowski Distance-based Centre Loss (MC Loss) and Minimum Margin Loss (MML) are presented in Chapter IV, and Precise Adjacent Margin Loss (PAM Loss) and Global Information-based Cosine Optimal s (GICO Loss) are presented in Chapter V. MC Loss extends the Centre Loss from the Euclidean distance to the Minkowski distance. MML enhances the discriminative ability of features by setting a minimum distance between all pairs of the class centres. PAM Loss penalises the margin and gives ‘margin’ a meaning that represents the real edge-to-edge margin between different classes in the training set. GICO Loss uses global information as the feedback information to optimise the intra-class and inter-class variance. To enable the calculation of GICO Loss, an algorithm is proposed to learn the cosine similarity between the class centre and the class edge. We conduct extensive experiments to evaluate these loss functions. Experimental results demonstrate their state-of-the-art performance. Furthermore, a complete face search framework for image/video has been proposed and evaluated in Chapter VI.
Date of AwardApr 2020
Original languageEnglish
SupervisorBryan Scotney (Supervisor) & Hui Wang (Supervisor)

Keywords

  • Image representation
  • Feature extraction
  • CNNs
  • Loss function
  • Face search

Cite this

Face Recognition Using Feature Fusion and Deep Learning
Wei, X. (Author). Apr 2020

Student thesis: Doctoral Thesis