Selecting an informative features vocabulary for recognition algorithms based on Fourier-descriptors
Vasily Kolyuchkin, Kong Nguen · Science and Education of the Bauman MSTU · 2014
Working vocabulary of features include most informative features of objects to be recognized. The aim is to develop a method of forming a working vocabulary of features for recognition algorithms based on Fourier-descriptors of the object image contours. To solve this problem the paper offers to use the method of functional maximization that is the ratio of the distance between the classes to the spread of objects within each of the classes represented in the feature space, which is formed on the basis of Fourier-descriptors. To check the effectiveness of the proposed method to form a working vocabulary of features the numerical experiments have been carried out. The experiments used two databases of reference images consisting of 10 and 13 reference images. Test images obtained by rotating the reference images, by zooming, as well as by adding the noise using the normal law of distribution have been created from these images. The proposed by the author algorithm, which uses the Prewitt operator, threshold segmentation, and morphological processing has marked the contours of images. The original vocabulary of features derived from the Fourier-descriptors has dimension of 98. The vocabularies of working features having the dimensions, respectively, 3 and 4 have been formed on the basis of functional maximization for both reference images. In the course of numerical experiments the frequency of correct decisions to recognise the features of reference bases of images for the original and working vocabularies has been evaluated. It has been proved that the algorithm of recognition with the formed working vocabularies of features provides a great efficiency of automatic recognition of objects. There are known publications, which use a similar method to form a working vocabulary of features in algorithms of human recognition by the image. But there are no publications on choosing the vocabulary of features for recognition algorithms based on the analysis of the image contours that can be used in computer vision systems of automated production lines.