New Parametric Statistical Models And Their Applications

Parametric statistical model is one of the most important feature extraction methods in computer vision and image understanding. However, classical parametric statistical models often extract image features merely from the gray intensities of an image, and omit the shape and geometric information of the objects in the image. Recently, new parametric statistical models, such as the active shape model and active appearance model, have been successfully used in the research areas as stated above and aroused extensive attentions. This thesis focuses on new parametric statistical models in one of the most important topics of image understanding, i.e. face image understating. In addition, the thesis considers the problems of face feature point localization and feature extraction in face image understanding along with other state-of-theart algorithms in computer vision and digital image understanding.Face image understanding is a technology that uses a computer to interpret the faces in an image or a video and subsequently to extract their high-level semantic meanings. As one of the most important components of digital image analysis and understanding, face image analysis is drawing extensive attentions in the communities of computer vision, pattern recognition, machine intelligence and cognitive psychology. The human face is one of the most vital organs of the human body. It plays an important role in people’s social activities and communications, and conveys very important interactive information. As a face image usually contains informative and discriminative features without physical contact, face image understanding has been widely used in biometrics, video surveillance, human computer interaction and digital multimedia. Many applications that emerge from face analysis have also been applied to various aspects in our daily social life.Face image understanding consists of a set of algorithms, including face detection, facial feature point localization, face recognition and classification. Much progress has been made in face image understanding during the past few years, and existing algorithms perform very well under controlled scenarios. However, due to the varieties in appearance and shape caused by illumination, expression and pose, a robust, efficient and effective face image understating algorithm is non-trivial and many problems need to be addressed. In face image understanding, statistical models have been widely used, such as for face feature point localization, feature extraction and classification. However, the classical statistical models assume that a face image has already been well aligned, and directly extract texture features without considering the problem caused by the natural variations of human faces. To deal with this problem, recently,active shape model and active appearance model have been widely used. For face image understanding, these new parametric models can not only extract effective facial texture features, but also obtain accurate shape and geometric information of a human face, which leads to a more effective face representation with semantic meanings.The thesis presents the advantages and characteristics of both parametric statistic models and non-parametric regression models, and proposes a set of new algorithms. We believe that these new findings will greatly enrich the face image understanding community. The main contributions of this thesis include:Firstly, we propose a tensor-based active appearance model built from an incomplete dataset with large-scale pose variations. The variations in appearance and morphology are two of the most important difficulties in face image understanding. In addition to that, two face images of the same identity can be very different due to variations in pose, expression, illumination and occlusion, which brings a great number of difficulties to facial feature extraction and analysis. Recently, to address this problem, tensor subspace analysis, also known as multilinear subspace analysis, has been successfully used in face image understanding. However, a large number of facial images are usually required to build a tensor-based model. To deal with this problem, this paper proposes to build a tensor-based active appearance model from an incomplete training data with missing values. Also, to address the problem of the inconsistency in shape and global appearance representation of faces with a large pose rotations, we present a unified facial landmarking strategy. The experimental results on the CMU MultiPIE face database demonstrate that the proposed algorithm performs well even when 80% of the training samples are missing, which validates the proposed algorithm.Secondly, we propose an adaptive non-parametric cascaded regression model for face feature point localization. Despite the success of active shape models, active appearance models and their extensions in controlled scenarios, these algorithms often fail in uncontrolled conditions. In recent years, the cascaded-regression-based methods have been successfully used in face feature point localization. However, these algorithms cannot always accurately localize face feature points due to the complexity of the appearance of a human face. To further improve a cascaded-regression-based model, this thesis proposes a random cascaded regression copse that builds a set of cascaded regression models in parallel by randomly selecting subsets from all the training samples. Moreover, to address the problem posed by scale variations, this part of the thesis presents an adaptive shape update strategy with an adaptive sparse auto-encoder based local feature extraction method. The experiments on LFPW and COFW show the effectiveness of the proposed method.Lastly, to perform data augmentation for a cascaded-regression-based method, we synthesize a large amount of training samples using a parametric statistical 3D morphable face model. The synthesized face images from a 3D face model are used to train a cascadedregression-based model along with a small number of real images. To adapt the trained model from synthesized faces to real ones, this thesis also proposes a two-step cascaded regression training strategy. The experimental results on a set of face datasets show the superiority of the proposed algorithm.