|Institution:||Nanyang Technological University|
|Keywords:||DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics|
|Full text PDF:||http://hdl.handle.net/10356/59211|
Facial image analysis contains many applications including facial landmark localization, face detection, facial expression recognition, gender recognition, eye open/close state recognition and others. Novel algorithms within two types of techniques: correlation ltering and feature extraction have been developed in this work based on these applications. Correlation ltering is one attractive technique to detect patterns of interest from images. This work presents an introduction of design theories for correlation lters, and proposes a new class of lters called Generalized Unconstrained Filters (GUF). Compared to existing lters, merits of GUF include: free of over- tting problem, and relief from the restriction that targets should be placed in the center of training images. More importantly, GUF provides an insight to unify many existing lters including Average Filter (AVG), Unconstrained Minimum Average Correlation Energy (UMACE), Maximum Average Correlation Height (MACH), Unconstrained Optimal Tradeo Synthetic Discriminant Functions (UOTSDF), Average of Synthetic Exact Filters (ASEF) and Minimum Output Sum of Squared Error (MOSSE). These lters are just special cases of GUF with appropriate parameters. Performance evaluation of GUF is conducted under two di erent scenarios. In one scenario, GUF lters are o ine trained using thousands of face images, and then applied to nd facial landmarks. In the other scenario, GUF is used as an online visual tracker, and provides a real-time solution to face detection under dif cult conditions by incorporating with the well known Viola Jones detector. For both scenarios, extensive experiments have been carried out to show the promising performance of GUF lters. On the other hand, feature extraction is an essential component of facial image analysis. Several feature extraction methods are presented to improve classi cation accuracy. First, a texture feature named Local Intensity Increasing Patterns (LIIP) is proposed to represent the intensity increasing trend within local intensity region. In theory, LIIP inherits the strengths from many existing approaches including Local Binary Patterns (LBP), uniform patterns and gradient information, and it represents local texture micro-structures in images. In addition, a variant of LIIP, the Local Gradient Increasing Patterns (LGIP) has been introduced to enhance the feature stability in case of white noise or non-monotonic intensity variations. Moreover, Multiscale Cell LIIP (MC-LIIP) extends original LIIP to scalable levels, so both texture micro-structures and macro-structures could be well described. Using these features, di erent strategies are adopted to extract feature vectors from face (big) and eye (small) images. Experiments show the proposed methods can achieve satisfactory performance in facial expression recognition, gender recognition and eye state recognition.