Towards a face recognition system : face detection, face registration, and head pose estimation

• Main Author: Ying, Ying
• Other Authors: Wang Han
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems; DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: http://hdl.handle.net/10356/59946
• Institution: Nanyang Technological University

Generally, solutions of face recognition system involve a series of computer vision topics with configurations and connections according to various requirements and emphases. In this dissertation, towards a face recognition system with no cooperation of human beings, face detection, face registration, and head pose estimation are fully and in-depth investigated, especially in the direction of being implemented as practical applications. It is with the aim to free these applications from the flaws in terms of accuracy, robustness, and efficiency that novel algorithms are proposed. Given any 2D image as input, the target of the face detection work is to determine face location and corresponding face extent as output, which is not only applicable to frontal and upright face view. Hence, a face detection system with embedded Statistic-based Subspace Analysis (SSA) is proposed for multi-view face detection, which is composed of four functional modules, preprocessing, coarse face detector, fine face detector, and postprocessing. The SSA method leads to an algorithm constructed by a hybrid subspace feature and a statistic-based discriminant function; the former uses local representation to reduce the sensitivity to image variances and describes characteristics comprehensively with the aid of multiple subspace features, while the latter determines final class label with the collections from all the statistic projections. For accuracy, the system is strengthened by two kinds of progressive structures; the vertical structure distinguishes face from false alarm by using the fine face detector learnt by the SSA method, and the horizontal structure separates pose range into a series of subranges, which helps to concentrate the discrimination ability of both the coarse and fine face detectors. For robustness, the system is invariant to complex background, illuminance, geometry, facial expression, and partial occlusion, owing to the discrimination ability of various features and classifiers in the coarse and fine face detectors. For efficiency, the system exhibits fully automatic working manner and real-time execution capability in respect that the vertical progressive structure filters most of the non-face regions by using the coarse face detector. The objective of the face registration work is to have relatively complete knowledge of face by determining the geometry relationship between pairwise 3D face data captured from the same subject and aligning them subsequently. As a result, a face registration system is presented, together with data association strategies and optimization mechanism. The Torchlight-Based (TB) data association strategy builds point correspondence with the constraint of virtual cylinder shape, and the Grid-Based (GB) data association strategy utilizes virtual cube shape instead; both strategies are capable of registering data with low quality and low overlap rate. With the aid of multi-level structure, the Adaptive Genetic Algorithm (AGA) approaches the final transformation result by taking account of both global and local optimizations, and with no requirement of prior knowledge. The characteristics of accuracy and robustness are guaranteed by two aspects. Firstly, the data association strategies, especially the TB strategy, remove non-overlapped data before building point correspondence and reduce the influence of interference by using the virtual shape constraint; secondly, the optimization mechanism narrows the geometrical space level by level and operates each evolution from both global and local points of view. The characteristic of efficiency is performed by the GB data association strategy which simplifies the calculation of virtual shape, and by the fully automatic working manner which adaptively updates space boundary by applying the AGA method. The head pose estimation work aims at inferring the head orientation relative to the real world coordinate by analyzing the depth information of face data. Therefore, a head pose estimation system, the Dynamic Random Regression Forests (DRRF), is proposed with four improvements distributing in the offline training procedure and online testing procedure. The offline training procedure includes three improvements: the boosting strategy for data induction assigns higher probability of being chosen to the error-prone data for pattern analysis, the dynamic binary test hierarchically optimizes the key parameters of classifier, and the stem operator increases the possibility of optimal data split. The improvement in the online testing procedure is the weighted voting scheme that intermediate outcomes vote for the final output according to their various voting reliability. The system demonstrates accurate performance for two reasons: the overall forest classifier is composed of a good many tree classifiers, each of which possesses considerable discrimination ability, and the four improvements are well designed throughout the data preparation phrase, classifier construction phrase, and data evaluation phrase. The system is robust against complex background, illumination variation, facial expression, and partial occlusion, owing to the randomness introduced during the offline training procedure. The system is also efficient, not only for the fully automatic working manner and real-time execution capability, but also for the additional face location information simultaneously obtained during the estimation of head pose.