Face class code based feature extraction for face recognition

In face recognition, the goal is to assign a class label for a test image of a subject from N classes in the database, when binary classifiers are used, the commonly used method is the one-per-class (OPC) i.e., one classifier per subject. A drawback of the OPC method is that when the number of classes is large, it takes very long time to make a classification decision. In place of the computationally-demanding OPC method, we propose a new feature extraction method "face class code" (FCC) based on binary classifiers. For example, correlation filters and support vector machines can be used to generate feature vectors to deal with large number of classes. The FCC method encodes each class label into a binary string, and we design classifiers to discriminate \´1\´ or \´0\´ for each bit in the sequence, to determine the class label. Thus, we will need as few as [log₂(N)] binary classifiers to achieve an N-class recognition problem. This binary coding framework also opens the whole world of error control codes (ECC), which can be used to improve the recognition performance. The proposed method is verified through experiments on the PIE database and the AR database.