Extraction of Illumination Invariants Using NSCT and Adaptive Smoothing

To better handle light changes in face recognition, a method combining NSCT and adaptive smoothing is proposed, to extract illumination invariants with more structural information of faces.The method first decomposes a face image in the logarithmic domain, then applies NormalShrink filtering to each high-frequency subband.Next, a blurred face image is obtained by performing inverse NSCT on the filtered high-frequency and original low-frequency subbands.After that, adaptive smoothing is used to extract detailed facial information from the original low-frequency subbands.Finally, a facial illumination invariant is estimated by combining the detailed facial information and the blurred face image.The resulting illumination invariant has more detailed facial information in the lowfrequency subbands, compared with results from NSCT.The proposed method better exploits the structural information of faces.Experiments on Yale B and CMU PIE face databases show that the method effectively eliminates illumination variations, has better performance compared with existing methods, and improves the robustness of face recognition systems.