DCT based region log-tiedrank covariance matrices for face recognition

摘要: Gabor-based region covariance matrix (GRCM) has been demonstrated as a promising descriptor for face recognition. However, GRCM requires large number of filters to achieve satisfactory performance. Furthermore, complex-valued Gabor require double convolution operations each filter that makes the computation more expensive. To alleviate problem, we propose adopt real-valued discrete cosine transform (DCT) bank in place filter. DCT an orthogonal however decorrelates signal, leads most energies fall into diagonal entries constructed matrix, which is ill-formed RCM. We demonstrate applying non-linear operation on responses ameliorates decorrelated effects. Apart from that, while RCM offers spatial information useful recognition tasks, overly small renders poor estimation, can affect performance drastically. In this paper also Log-TiedRank mitigate potential undersampling effect suffered by estimation. From experiments shows surprising boost over AIRM and Log-Euclidean metric especially when both gallery set probe have very different distributions.