Optimal discrete wavelet transform (DWT) features for face recognition

Face recognition systems usually include preprocessing, in order to crop the training and probe images. This often involves arbitrarily-chosen segmentation boundaries, which may exclude discriminative face information or include irrelevant pixels corresponding to background, hair, etc. The work presented in this paper creates a rich feature vector using discrete wavelet transform (DWT) coefficients, which is then optimized to exclude useless information. This optimization process eliminates the need to overly crop images, as background will be automatically excluded. Experiments on the AT&T database show that the technique improves results significantly, with recognition rates increasing from 93% to 97.5% when using the Haar wavelet.