Semester Topic: Acquisition, Representation and Processing of Information
Group Members: Carsten Høilund, Jeppe Jensen, Thomas Paulin, Simon J.K. Pedersen
Abstract:
This 6th semester report concerns basic pattern recognition with special attention on face recognition. The objective is to utilize face recognition in a photo album context to enable automatic categorization of images according to the person present in the images.

The starting point of the face recognition system is the general pattern recognition system. The first step is preprocessing and segmentation, which involves extraction of the faces in images similar to passport photos. Features are extracted from the face images, using Principal Component Analysis (PCA). PCA determines the eigenfaces of the face images, after which the weight vectors corresponding to the most significant eigenfaces are used as features for the classifiers.

The primary focus of interest is the pattern classification techniques. For this reason four different classifiers are analyzed and evaluated with respect to their usability in a photo album. The four methods are:

• Agglomerative Hierarchical Clustering (AHC)
• Bayesian Decision Theory
• Linear Discriminant Functions (LDF)
• k-Nearest-Neighbor (k-NN)

The classifiers are i.a. tested for three common factors; the minimum number of training samples needed to obtain an acceptable recognition rate, the minimum resolution of the face images usable, and the computation time. The tests showed that k-NN delivered the best recognition rate based on the criteria of fewest training samples. Thus the most suitable of the investigated classifers for a photo album is k-NN.