Face recognition is an important field of research with many potential applications for suitably efficient systems, including biometric security and searching large face databases. This paper describes an approach to the problem based on a new type of n-tuple classifier: the continuous n-tuple system. Results indicate that the new method is faster and more accurate than previous methods reported in the literature on the widely used Olivetti Research Laboratories (ORL) face database.
Conventional n-tuple systems have the following desirable features:
In conventional n-tuple based image recognition systems, the locations specified by each n-tuple are used to identify an address in a look-up-table. The contents of this address either use a single bit to indicate whether or not this address was accessed during training, or store a count of how many times that address occurred. While conventional n-tuple methods are very fast, both for training and recognition, the size of the address space quickly becomes excessively large if the values at each pixel location in the image are grey scale rather than binary, and while a number of methods have been proposed to overcome this, none of them seem entirely satisfactory. The approach adopted here is to store them sparsely i.e. rather than store the whole address space in memory, just store explicitly the set of vectors that have occurred for each n-tuple. Each vector is tagged with its class. Then, during testing, the score for each n-tuple for each class is then simply the distance to the nearest vector of that class. The overall score for each class is the sum over all n-tuples of each n-tuple score for that class. The pattern is then classified as belonging to the class with the lowest overall score.
The continuous n-tuple system offers greater accuracy, while retaining most of the advantages of the conventional n-tuple system. There is, however, a tradeoff as will be seen below, between the recognition accuracy, the training time, and the recognition speed. In the case of the face recognition problem described below, continuous n-tuple systems have been created that are more than competitive with previous approaches.
The rest of this paper is structured as follows: the next section describes the conventional n-tuple classifier, section 3 describes the continuous n-tuple classifier, section 4 presents results on the ORL face database, and section 5 concludes.
Adrian F Clark