Automatska klasifikacija slika zasnovana na fuziji deskriptora i nadgledanom mašinskom učenju

Abstract

This thesis investigates possibilities for fusion, i.e. combining of different types of image descriptors, in order to improve accuracy and efficiency of image classification. Broad range of techniques for fusion of color and texture descriptors were analyzed, belonging to two approaches – early fusion and late fusion. Early fusion approach combines descriptors during the extraction phase, while late fusion is based on combining of classification results of independent classifiers. An efficient algorithm for extraction of a compact image descriptor based on early fusion of texture and color information, is proposed in the thesis. Experimental evaluation of the algorithm demonstrated a good compromise between efficiency and accuracy of classification results. Research on the late fusion approach was focused on artificial neural networks and a recently introduced algorithm for extremly fast training of neural networks denoted as Extreme Learning Machines - ELM. Main disadvantages of ELM are insufficient stability and limited accuracy of results. To overcome these problems, a technique for combining results of multiple ELM-s into a single classifier is proposed, based on probability sum rules. The created ensemble of ELM-s has demonstrated significiant improvement of accuracy and stability of results, compared with an individual ELM. In order to additionaly improve classification accuracy, a novel hierarchical method for late fusion of multiple complementary descriptors by using ELM classifiers, is proposed in the thesis. In the first phase of the proposed method, a separate ensemble of ELM classifiers is trained for every single descriptor. In the second phase, an additional ELM-based classifier is introduced to learn the optimal combination of descriptors for every category. This approach enables a system to choose those descriptors which are the most representative for every category. Comparative evaluation over several benchmark datasets, has demonstrated highly accurate classification results, comparable to the state-of-the-art methods.