In this paper a new method is being proposed which makes use of symbolic data manifested by empirical cumulative distribution functions (ECDF) and distribution functions based on sets of ECDF values, referred to as the distribution function of distribution values (DFDV) of image features. This differs with conventional image classification studies, which mostly rely on pixel intensity values as features. Such symbolic representations provide a more general characterization of the pixel intensities patterns across image areas. The main novelty associated with the given research is the creation of the DFDV based on the best possible selection of the distinguishing points in order to capture fundamental changes in intensity distributions within image classes. Compared to previous studies which have used pre-determined distinguishing points, the proposed study proposes a clustering-based approach to find distinguishing points that maximise class separability. Experimental evaluation on the MNIST handwritten digits data set demonstrates the effectiveness of the proposed method, achieving an average classification accuracy of 68.27% and a highest accuracy of 95.33%. These results indicate that integrating clustering-based symbolic features extraction with copula-based modelling provides a competitive and promising for image classification tasks.
Winarni et al. (Mon,) studied this question.