Abstract: Face recognition has attracted much attention because of its wide applications. However, face recognition is still an unsolved problem as human face is not a rigid object and it can be transformed easily under large intra-class facial variations such as pose, illumination, expression and small inter-class difference. In some cases the difference of face image of same person could be larger than those from different one. Therefore, how to represent the intrinsic attributes of a human face effectively becomes much more important to increase the accuracy of face recognition systems. Proposed system uses person-specific SIFT features and a simple non-statistical matching strategy to solve face recognition problems and also uses t-test algorithm for feature selection. First, Region of interest is deduced from the face image using object detector which has advantage over Haar cascade object detection which is effective only on frontal image. In the feature extraction step, Scale Invariant Feature Transform (SIFT) is applied. In contrast to Gabor features, SIFT features are invariant to image scaling, rotation and also partially invariant to illumination. Then the dimension of the feature space is reduced by defining features over regions of interest that are selected by t-test feature selection with feature correlation weighting. T-test is done based on high probability feature index with common features from different class is given low probability value and also high probability value for discriminant features that is enlarge the variation between the classes and minimize the similarity .In image classification, matching strategy is used. If the matching value is high the image is recognized. Experimental results demonstrate that the proposed algorithm yield superior performance with much lower dimensionality as compared to performance on the original data or on data transformed with other dimensionality-reduction approaches.
Keywords: Scale Invariant Feature Transform, T-test.