Face Recognition Based On Dual-tree Complex Wavelet Feature

Automatic Face Recognition technique is a type of Biometric Authentication technique for identifying humans, in which face images are processed and analyzed by computer techniques to exact effective recognition information. Face Recognition is a typical pattern recognition problem, it not only depends on the design of classification machine, but also depends on the process of exacting features. Approximate shift invariance, good directional selectivity, computational efficiency properties of Dual-Tree Complex Wavelet Transform (DTCWT) make it a good candidate for representing the face features. 2D DTCWT can isolate edges with six orientations in different sub-bands, which avails to description of texture features in different orientations. The main purpose of my research is how to exact most discriminative and robust face features utilizing DTCWT. This paper proposes two novel face feature extraction methods based on DTCWT. One method is based on DTCWT magnitude and subspace methods, the other based on DTCWT phase and local histograms.For DTCWT magnitude and subspace based face representation, first, we combine DTCWT magnitude in multi-scale and multi-direction into a large 2D matrix, and down-sample it to the same size of original image. Then, its dimension is descended by linear subspace methods. Thus, we obtain the face representation based on DTCWT magnitude and subspace. DTCWT magnitude and subspace based face representation is robust to variation of lighting and expression. It outperforms Gabor magnitude based face representation, because of its more efficient computation and no need to select frequency parameters. The DCF method, which combines the advantage of DTCWT and Fisherface, obtains high recognition rate and good generalization ability both under different number of features and training samples.In Histogram of DTCWT phase based face representation method, for each sub-image of DTCWT, we first get Quadrant bit coding of phase, and encode the local neighborhood variations by a local XOR pattern (LXP) operator. Then what extracted is local spatial histograms of each sub-image. Finally, we concatenate these into an extended histogram feature to represent the original image. Quadrant bit coding of phase, local XOR pattern, local spatial histograms, which are adopted by this methods, together made the feature representation robust to variations of illumination lighting, expression and misalignment. In addition, this face modeling method does not need the training set for statistic learning, thus it avoids the generalization problem. Experimental results show that it can achieve high recognition rate, is an ideal face feature description method.