A Gabor Filterbank Approach for Face Recognition and Classification Using Hybrid Metric Learning

In this paper, we consider the notion of distance/similarity metric learning method for classification and similarity search for face recognition purposes. We use the theory of filter banks, and Gabor wavelets for extraction of face features in three datasets: the ORL dataset, Yale face dataset, and ATT face dataset, and we compare the recognition performance with Gabor features to discrete Fourier transform baseline method for feature extraction. For the learning application part, there are various different approaches proposed under the distance metric learning framework. We base our work on two different approaches: i) Distance metric learning for LMNN [1], ii) Information Theoretical Metric Learning (ITML) [2], and. Our main goal is, employing optimization-based techniques, to combine two distance metrics into a single hybrid metric explaining the similarity/dissimilarity relation better than each of its base metrics. Since the selection of features impacts the classification and recognition results, we propose that Gabor features obtained using a parallel bank of Gabor filter at different scales and orientations provides the best recognition rate. Furthermore, we attempt to perform the objective of collapsing two distance/similarity metrics into a single one by optimizing an appropriate weighting which enables the hybrid metric to inherit the useful characteristics of its component metrics in terms of classification accuracy with similarity/distance information. We conduct experiments with hybrid metric approach on various datasets to compare its performance to that of each of its components individually. To this end, we first provide a general background overview on Gabor filters for feature extraction, distance metrics and some special family of convex optimization problems along with the metric learning approaches introduced in [1], [2], and [3]. Finally, we introduce our hybrid metric learning approach and present our experimental results obtained from the face data-sets.