Classifying images using multiple binary-class decision trees for object-based image retrieval

This paper describes an approach to multiclass object classification using local information based invariant object-contour representation and a combination of one-per-class binary-class decision tree classifiers. The object representation scheme is based on the polygonal approximations of object contours. C4.5 is used to learn each of the binary-class tree classifiers which are used to predict the class of each segment of an object. A new decision combination method is used to determine the class of an object based on class probability distribution of each segment of the object on each of the binary-class trees. The proposed object classification approach is invariant to translation, rotation, and scale changes of objects. On applying this approach to a hand tool image database in the situation of image retrieval, the experimental results show that the retrieval performance is significantly better than the results obtained by previous studies.