Undecimated dual tree complex wavelet transform based face recognition

摘要

In this paper, we have developed a local descriptor and two global descriptors based on the Undecimated Dual Tree Complex Wavelet Transform (UDTCWT). Undecimated dual tree complex wavelet transform possesses certain advantages over the traditional wavelet transforms and hence it is capable of representing digital image signals more accurately. We have explored this concept considering the face recognition problem. Given a face image we compute the complex UDTCWT coefficient images of the face image at 4 scales and 6 orientations. Using these coefficient images we compute 48 Local UDTCWT Phase Patterns (LUPPs) and 8 Global UDTCWT Phase Patterns (GUPPs). Dividing these patterns into blocks and concatenating the 2D magnitude weighted phase histogram of the complex coefficients in these blocks, we form our Global descriptor. To handle pose and expression variation in face images, we have developed a key point based local descriptor. Given a face image, using the box filter response scale space, we have obtained scale dependent size square regions around interest points and these square regions are represented using UDTCWT. Extensive experiments conducted on benchmark face recognition datasets FERET, ORL, YALE and UMIST have demonstrated the appropriateness of our descriptors for face recognition applications.