On the solution of the image recognition problem by a principal component method and linear discriminant analysis

In the paper, some aspects of image analysis based on PCA (Principal Component Analysis) and LDA (Linear Discriminant Analysis) are considered. The main idea of this approach is that, firstly, we project the face image from the original vector space to a face subspace via PCA, secondly, we use LDA to obtain a linear classifier. In the paper, the efficiency of application of PCA and LDA to a problem of recognition of face images without their preliminary normalization is investigated. When the number of images in a class is not large, it is proposed that the training set is supplemented by images obtained by rotating, scaling and mirroring. In the images from the ORL and Feret databases, the influence of the training set expansion on the quality of recognition of unnormalized face images is studied. Also, a problem of increasing the efficiency of principal component calculation for large image samples is addressed. A linear condensation method is used as a new technique to calculate the principal components of a large matrix. The accuracy and performance of the developed algorithm are evaluated.