ISSN 1054-6618, Pattern Recognition and Image Analysis, 2015, Vol. 25, No. 4, pp. 692–704. © Pleiades Publishing, Ltd., 2015. 1 1. INTRODUCTION This work is devoted to the development of a fea- ture description of images of paintings for attribution purposes. By the term “attribution” is meant the assessment of whether an unauthorized piece of art belongs to one or other author, school, time, country, etc. [1, 2]. In the recent decade, a large number of investiga- tions of the works of pictorial art have been carried out by the methods of computer-aided analysis of images [3, 4]. One of directions was associated with obtaining quantitative characteristics of the artistic manner of painters and the development of methods for comput- erized attribution [3]. These methods are based on the comparison of the quantitative characteristics of a painting of unknown attribution with the characteris- tics of authentic works. Analysis of publications shows that the solution of the problem of computerized attribution involves sev- eral main stages. The scheme of the solution is shown in Fig. 1. At the stage of preliminary processing, dis- tortions arising in digitized images during acquisition are compensated for, and the images are scaled to obtain identical resolution [5, 6]. Features are calcu- lated on relatively small fragments of paintings. There- 1 This paper uses the materials of the report submitted at the 11th International Conference “Pattern Recognition and Image Analysis: New Information Technologies,” Samara, Russia, September 23–28, 2013. fore, an image under test and an image of known attri- bution are divided into fragments, from which features are extracted. At the next stage, one calculates the val- ues of local (within a fragment) features. In many studies, the authors use the global features of an image of a painting. The dimension of the feature space may be large enough. Therefore, at the next stage, one car- ries out the aggregation of features within individual fragments to reduce the dimension [6]. Various stan- dard methods have been applied to reduce the dimen- sion of the feature space [5, 7, 8]. After reducing the dimension, the fragments of images under test are compared with the fragments of images of authentic paintings by the chosen similarity or dissimilarity measure. The distance between paint- ings is obtained as a result of aggregation of distances between fragments. If the images of paintings are not defragmented, the distances are calculated directly by extracted feature vectors. Then, on the basis of dis- tances obtained, one makes attributional decisions. Various research groups have proposed the follow- ing approaches to the solution of the problem of com- puterized attribution. The first approach is based on the enumerative comparison of square fragments into which both images under test are decomposed [5]. As a rule, features are given by the coefficients of orthog- onal transformations (in particular, wavelet transfor- mations). These methods require large computational resources and are very sensitive to the conditions of image acquisition and to the parameters of hardware [9].The second approach [6, 10, 21–23] is associated with the localization of specific objects in images by which the images are compared. These objects are Feature Description of Informative Fragments in the Problem of Computerized Attribution of Paintings 1 D. M. Murashov Dorodnitsyn Computing Centre, Russian Academy of Sciences, ul. Vavilova 40, Moscow, 119333 Russia e-mail: d_murashov@mail.ru Abstract—Problems of composing a feature description and developing a comparison procedure for the images of paintings in attribution are considered. A feature description of a facture of paintings based on the characteristics of a grayscale image relief and elements of the structure tensor is proposed. In contrast to known techniques, the feature description is formed only by informative fragments of images and does not require preliminary segmentation of individual brushstrokes. Parameters of feature extraction procedures are chosen. Measures of dissimilarity between images of paintings are proposed. Computational experiments are carried out. The feature description proposed is a quantitative characteristic of the artistic style of an author. The procedure developed for comparing images can be applied together with other types of investigation of paintings to make an attributional conclusion. Keywords: feature description, facture of paintings, attribution, artistic manner, ridges of an image, structure tensor, measure of dissimilarity of images, Kullback–Leibler divergence, images of paintings. DOI: 10.1134/S1054661815040197 Received March 14, 2014 APPLIED PROBLEMS