Fusion Of Gabor And Two-dimensional Subclass Discriminant Analysis For Face Recognition

Face recognition is a very important research direciton in the pattern recognition field. Overthe past decades, a large number of automatic face recognition algorithms play an important rolein application domains, such as visual surveillance, access control and identity informationverification. Face recogniton problem involves two aspects: face detection and feature extraction.This paper focuse on the development of face recognition and introduces the current popularfeature extraction algorithms which can ectract characters from human faces accurately. Then thispaper proposes a fusion of Gabor and Two-Dimensional subclass discriminant analysis for facerecognition.The discriminant analysis(DA) is the most important feature extraction methods incomputer vision applications. This algorithms has main advantage is that they can extract alow-dimensional feature from high-dimensional space,while the feature can be effectivelyseparated by their face class samples, such as PCA,. Linear Discriminant Analysis (LDA). Themain idea of LDA algorithm aims to achieve the optimal classification by the best projectiondirection. It is based on fisher principle that maximize the between-class distance whileminimizing the within-class distance. Unfortunately, whether the principal component analysismethod PCA or linear discriminant analysis LDA algorithms have some deficiencies. LDA hassome restrictions in applications. It assumes the class samples vaectors are normal distributions ofcommon covariance matix but different means. But in fact, the different of illumination andexpression can lead to facail image data non-convex complex distribution. So under the differentof illumination and expression conditions, LDA algorithm will decline the face recognitionaccuracy. This is one of the important issues in face recognition algorithms.This paper proposes a fusion of Gabor and Two-Dimensional subclass discriminant analysisfor face recognition. This algorithm aim to sovle the decline of face recognition accuracy underdifferent of illumination and expression conditions. Fist, this algorithm decompose the face imageby, then convole the face image with Gabor filters in multi-orientation and multi-scale. But afterGabor wavelet the face samples will produce a large number of samples, so the samples need tobe downsampling.2DSDA algorithm is used to decrease the dimension of the Gabor sample. Ithas two advantages:1.2DSDA can advoid small sample size problem and use the structure offace data.2. It can divide face samples belonging to the same class. Combined with within-class and inter-class eigenvector matrix,. this paper suppose a adaptive stability criterion to divide thesame samples. Finally the nearest neighbor rule is used to classify.The experiments on ORL and AR face database show that the proposed algorithm has higerrecogniton accuracy and be better robust to the different of illumination and expression, and fromthe experimental results, it can see that the algorithm has higher stablility then other algorithm.The experiment proves the accuracy and practical value of the Gabor-2DSDA algorithm.