Appearance based face recognition system with mobility

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Lisa Osadciw

Second Advisor

Pinyuen Chen


Face recognition, Learning curve, Fusion, Wireless, Confidence interval

Subject Categories

Electrical and Computer Engineering


Face Recognition, as a non-intrusive and non-contact biometrics, has been gradually accepted by the general public to be used in authentication, forensics, and security. The active research and commercial efforts keep raising the bar of recognition accuracy while dragging down the equipment costs. However, the scalability of such a system is still a challenging problem. Motivated by the learning curve phenomenon, we propose intelligent sampling to construct a representative and efficient training database. Initial clustering is implemented to measure the difficulties of the training images to the current system, and the recognition performance is evaluated to determine whether and what training images should be included. Intelligent sampling alleviates the overfitting problem. This training set construction minimizes training yet takes advantage of the full potential of face recognition algorithms without a complete system overhaul. After the features are extracted, a modified indifference zone method (MIZM) is proposed for feature selection. It is shown that the objective function of MIZM is consistent with the Linear Discriminant Analysis (LDA) method. MIZM extends the ranking and selection procedure from Principle Component Analysis (PCA) to LDA, supporting better face recognition performance than using the Scree plot or using all images. Features are further screened by an unsorted discrete particle swarm optimization algorithm. A product rule is proposed to update the likeliness, which double-asserts the selections of several attractors of the particle. A validation set is utilized for selection, and a disjoint test set proves the generalizability of this method on unseen images. Besides feature extraction and selection, diversity in feature extraction helps to improve face recognition performance through similarity score evaluation, and multiple classifier decision making. The recognition performance, from the fusion of a strong classifier and several weak classifiers at score and decision levels, improves from either classifier alone. It is illustrated that uni-modal face recognition fusion is effective. At last, motivated by the demand of mobile identification units and the prevalent availability of wireless networks, we propose a distributed wireless face recognition system. Image compressions, distributed calculations, and the effect of transmission errors on the final recognition performance are studied.


ISBN 9781109450705