Face Recognition Based On Improved PCA-class Algorithm

Human face recognition technology has very high academic and practicalvalues. Automatic facial recognition has been a longstanding challenge in the fieldof computer vision and pattern recognition for several decades. Feature extraction isone of the central issues for face recognition. Subspace learning is often used as afeature extraction method. The idea of subspace learning is to project the featurefrom the original high dimensional space to a low dimensional subspace, which iscalled projective subspace. In the projective subspace, the projected feature is easierto be distinguished than the original feature. Principal component analysis(PCA) isone of the widely used linear subspace learning algorithms.The unbalanced power spectra of facial images will result in some potentialproblems in face recognition. In the thesis, a Whitening PCA-class algorithmframework based on the concept of whitened-faces is proposed. In the algorithmframework, first preprocess the original images by a whitening filter and a low-passfilter, then extract feature vectors (or matrices) combined with the traditionalPCA-class algorithms, and finally complete the face recognition through the k-NNclassification method. The experimental result on the ORL face image databaseshows that the algorithm framework brings better recognition performance andachieves higher recognition accuracy.A new approach called modular two-directional two-dimensional PCA(M(2D)~2PCA) is presented. The proposed technique is based on block-image andtwo-directional two-dimensional principal component analysis ((2D)2PCA). First,in the proposed approach, images are divided into some blocks. Then, the(2D)~2PCA method is used to the blocks obtained from the previous step. It aims tofurther improve the recognition rates for face images with large variation in light,pose and facial expression. Experimental results on ORL and UMIST face databaseshow that the proposed method can achieve same and even higher recognitionaccuracy than Modular2DPCA with lower dimensions.