Fast Active Tabu Search In Image Classification Applications

With the network resources becoming more diversified and intelligent, image as a form of multi-media has the advantages of specific, vivid and visually. The number of electronic image grows exponentially with the development of the Internet, which presents great demands for image processing. Classification is one of them. The most primitive method of hand-labeling occurred in 1970s, which needs much human labor. Then a lot of traditional methods of image classification emerged, in which included the methods based on the keywords and the methods based on the image itself. Classification based on image itself seems more intuitive, and many classification methods can be adopted, such as k-NN, decision tree, SVM and so on. But to achieve high accuracy is still quite challenging.This thesis presents a novel framework for image classification:The training step finds an optimal cycle in each class connecting similar images in order through minimizing the geometric manifold entropy using fast active tabu search. In the testing step, an unknown image is grouped into a class if the entropy increase as the result of inserting the image into the optimal cycle of this class is the least. The proposed approach can generalize well on difficult image classification problems, especially, for images taken in different views. Experimental results show that the proposed framework is feasible and has potential in the application of image-based modeling retrieval.