Hierarchical clustering on texture statistics for search space reduction of large scale face recognition

The complexity of the face recognition like deformable pattern recognition lies in the variation of captured face images in terms of pose, illumination and expression (PIE). Recognizing a face image, captured in unconstrained environment, from a gallery of large face database looses its computational efficiency in terms of speed rapidly. To segregate a suitable subset of the database from the whole gallery researchers have already proposed plenty of data-mining techniques [1], [2], [3]. The proposed method of data mining is an improvement over texture statistics based clustering [1] proposed earlier. In the present method we have given importance to the feature weights obtained from decision tree [2] instead of one shot clustering [1]. A step-by-step hierarchical clustering is proposed for fast convergence and score improvement. We have considered the higher order statistics of the feature texture instead of considering only second order moment [1], [2], for reducing the interclass overlapping region. We have tested our algorithm on facial images of CDAC and FERET face database by two different algorithm namely Elastic Bunch Graph Matching (EBGM) and Scale Invariant Feature Transform (SIFT).