Identification of Untrained Facial Image in Combined Global and Local Preserving Feature Space

In real time applications, biometric authentication has been widely regarded as the most foolproof - or at least the hardest to forge or spoof. Several research works on face recognition based on appearance, features like intensity, color, textures or shape have been done over the last decade. In those works, mostly the classification is achieved by using the similarity measurement techniques that find the minimum distance among the training and testing feature set. When presenting This leads to the wrong classification when presenting the untrained image or unknown image, since the classification process locates at least one wining cluster that having minimum distance or maximum variance among the existing clusters. But for the real time security related applications, these new facial image should be reported and the necessary action has to be taken accordingly. In this paper we propose the following two techniques for this purpose: i. Uses a threshold value calculated by finding the average of the minimum matching distances of the wrong classifications encountered during the training phase. ii. Uses the fact that the wrong classification increases the ratio of within-class distance and between-class distance. Experiments have been conducted using the ORL facial database and a fair comparison is made with these two techniques to show the efficiency of these techniques.