Hierarchical Intelligent Classification Systems

Abstract

This chapter is focused on the insufficient performance of the hierarchical classifiers. The theory of granular computing and the mathematical model of the piecewise-regular object from Chap. 2 are used to define the hierarchical levels. The sequence with a small number of the weakly homogeneous segments is considered as a coarse-grained granule. Fine-grained granules stand for a large number of highhomogeneous small segments. We apply sequential three-way decisions (TWD) to speed-up the classification procedure. To improve the classification performance at each granularity level, the probabilistic rough set of the distance between objects from different classes at each level is created. If the distance between an observed object and the next checked instance is included in its negative region, the search procedure is terminated. Experimental results demonstrated that sequential TWD significantly decreases the classification time in comparison with the matching of the pyramid histograms of oriented gradients. Granular Computing and Hierarchical Intelligent Classification SystemIt is known that the theory of granular computing [10,20] can be used to significantly improve the recognition quality of the audiovisual data by its examination at various granulation levels [5,7] in a hierarchical classification system [6]. It is the key idea of the CNN, which has been recently proved to reach the quality of human recognition in several image recognition tasks [3]. One of the most practically important examples of the potential of granular computing in the classification task is the PHOG (Pyramid HOG) image descriptor [1]. According to this method, an image is divided at each granularity level into a regular grid of blocks and the HOG is computed in each block. The pyramid of grids is built, and the HOGs of the input and reference objects are compared at each level of hierarchy (pyramid). The final distance between objects is calculated as a weighted sum of distances between the HOGs in the pyramid. It was shown [1] that this method allowed to achieve 10 %-higher accuracy in comparison with the state-of-the-art HOG [4]. Let us discuss the hierarchical classification of the piecewise-regular objects (Chap. 2) in detail.