ALLPCB
Overview
A commonly used regional feature analysis approach in face recognition combines computer image processing and biostatistics. It extracts facial feature points from video using image processing techniques and applies biostatistical methods to build a mathematical model, i.e., a facial feature template. The template is compared with a subject's facial image to produce a similarity score, which is used to determine whether the faces belong to the same person.
Main methods
(1) Geometric feature methods: Geometric features include the shapes of the eyes, nose, mouth, and their geometric relationships, such as distances between them. These algorithms are fast and require little memory, but their recognition accuracy is relatively low.
(2) PCA-based methods (eigenfaces): The eigenface method is based on the Karhunen–Loève transform, an optimal orthogonal transform for image compression. High-dimensional image space is transformed into a new set of orthogonal bases; retaining the most significant bases spans a low-dimensional linear space. If facial projections in this low-dimensional space are separable, these projections can be used as feature vectors for recognition. These methods require a relatively large number of training samples and rely entirely on statistical properties of image gray levels. Several variants of eigenface methods have been proposed.
(3) Neural network methods: Neural networks can take inputs such as reduced-resolution face images, local autocorrelation functions, or second-order moments of local texture. These methods also require many training samples, while in many applications the available sample size is limited.
(4) Elastic graph matching: Elastic graph matching defines a distance in 2D space that is invariant to common facial deformations and represents a face by an attributed topological graph. Each vertex of the topology contains a feature vector that records information near that vertex. This method combines gray-level features and geometric factors, allows elastic deformations during matching, and achieves good resistance to expression changes. It also does not require multiple training samples per person.
(5) Line-segment Hausdorff distance (LHD): Psychological studies show humans can recognize contour images as quickly and accurately as gray-level images. LHD is based on line-segment maps extracted from facial gray-level images and defines the distance between two line-segment sets. Unlike methods that establish one-to-one correspondences between segments, LHD adapts to small variations between line-segment maps. Experiments show LHD performs well under varying illumination and pose, but its performance degrades with large facial expressions.
(6) Support vector machines (SVM): SVMs are a statistical pattern recognition approach that seek a compromise between empirical risk and generalization ability to improve performance. SVMs mainly address binary classification by mapping a low-dimensional linearly inseparable problem into a high-dimensional linearly separable space. Experimental results often show good recognition rates, but SVMs require large training sets, which is impractical in many applications. Training can be time-consuming, implementation is complex, and there is no unified theory for choosing kernel functions.
