Abstract
In the supervised learning of classifiers, there are cases where the target is not properly segmented. These include the following: 1) the multiclass case: the target contains objects that belong to different classes; 2) the incomplete case: parts of the objects are missing from the target; and 3) the other-class case: the target contains objects that belong to classes other than those being considered. In order to handle the multiclass case, a classifying procedure is employed that dynamically creates merged classes. In order to train merged classes, it is required that each feature is defined on very small domain, and that the range of each feature is binary, i.e., {0, 1}. In fact, there is no contradict to consider the range of each feature is binary when the naive Bayes classifier is used in the bag-of-keypoints technique, and so, the multiclass case is handled by re-arranging bag-of-keypoints technique. The result is presented for the experiment that verifies the classifier correctly detects multiclass targets. Also, proposing technique is extended to ordinary Bayes classifiers, and it is discussed how multiclass case is solved by the support vector machine.
