Face Recognition Based On The Features Extracted In The Fractional Domain

Face recognition, as one of the most typical applications, has been widelyconcerned and applied in the field of information security, e-commerce, artificialintelligence and other fields. Traditional face recognition technology is mainly basedon gray image, which is the most familiar method and has the history of severaldecades, but its recognition accuracy in practical applications is still difficult to meetpeople’s expected requirements, especially in the presence of the images capturedillumination changes, camera orientation and other interference variation. Primitivefacial image in the face recognition system is typically expressed as a set of grayvalue pixel grid. Because the isolated set of pixel gray cannot directly reflect theintrinsic characteristic of the face images, the introduction of an appropriate transform,mapped to the transform domain to make the face recognition, is an effective way toimprove the performance of the face recognition. As the generalized form of theFourier transform, the two-dimensional discrete Fractional Fourier transform(2D-FRFT) can be interpreted as a rotation of signals in the time-frequency plane.Fractional Fourier Transforms of different orders of the face images, corresponding todifferent fractional domain space, can be more effective to extract the features tocharacterize the face images. This paper studied the properties of the fractionaldomain of the face images, and proposed the feature extraction methods, the mainwork and contributions of the paper are as follows:1. This paper proposed the amplitude domain of the Fractional FourierTransform based sparse representation and classification method (FRFT-SRC). Theuse of multi-order amplitude features of the Fraction domain, not only enriched thefeature domain of the face images, and can calculate a compression FRFT basedcondition dictionary. It can reduce the dimension of the conditions dictionary, andthen the computational complexity had been reduced. Meanwhile, based on theenergy aggregation of the amplitude features of the fractional domain, this paperproposed Sparse-PCA based feature extraction methods of the Fractional Fourier Transform.2. The study of the properties of the phase domain of the Fractional FourierTransform, it got that the phase components, which have the most important visualinformation and the innate properties of the face shape, are very important in facerecognition. Based on this, this paper proposed a texture feature extraction methodwhich was based on multi-order phase domain of the Fractional Fourier Transform ofthe face images and the local XOR pattern. Meanwhile, in order to reduce thedimension of the texture feature vector domain, and avoid the small sample sizeproblem (SSS) when performing feature extraction, this paper also proposed blockbased Fisher discriminant feature extraction method.3. To extract the rich, effective and complementary feature sets of the faceimages, three kinds of feature extraction models based on the amplitude domain, thereal and imaginary domain and the phase domain of the two-dimensional discretefractional Fourier transform are presented. Based on the different characteristics ofthe three features in nature, three complementary set of features are extracted, thencalculate three kinds of similarity matrix of the feature extraction model classifier.Finally, at the decision level, the three similarity matrices generated using the threefeature extraction model are fused using a weighted sum rule to derive the finalsimilarity matrix and then nearest neighbor classifier is used for classification to getthe recognition result. Compared with the face recognition method based on singlemode, the recognition performance can be significantly improved.