Description of images using model-oriented descriptors

The paper proposes an approach to constructing an image description using a set of model-oriented descriptors. Each descriptor characterizes the "similarity" of the analyzed image, represented as a complex-valued gradient field, to a pre-selected model of this descriptor. It is proposed that descriptor models should be synthesized using a method of principal components, or discriminant analysis, which has been applied to a diversity of complex-valued gradient field realizations. As a result, the proposed approach enables the complex-valued field of the gradient of the analyzed image to be described as a set of real quantities from the interval [0,1], capable of simultaneously characterizing the phase and magnitude of the image gradient. The effectiveness of the proposed approach is illustrated via solving a face recognition problem and comparing the result with prototype solutions (based on the principal component method and discriminant analysis), which directly utilize halftone images. The comparison is made using a nearest neighbor's classifier.