Fast Objects Detection by Variance-maximization Learning of Lifting Wavelet Filters

A fast objects detecting method is proposed, which is based on the variance-maximization learning of lifting dyadic wavelet filters. First, it is shown that the sum of lifting wavelet coefficients in horizontal and vertical directions defines an elliptic-type of discrete operator containing free parameters. The free parameters are learned so as to maximize the variance of lifting wavelet coefficients for a target object. Since this problem is an ill-posed problem, a regularization method is employed to solve it. Objects in a query image similar to the target object are detected by the use of the learned filter. Simulation concerns the detection of narrow eyes from faces.