A new progressive open-set recognition method with adaptive probability threshold

Abstract

In the traditional pattern classification method, it usually assumes that the object to be classified must lie in one of given (known) classes of the training data set. However, the training data set may not contain the class of some objects in practice, and this is considered as an Open-Set Recognition (OSR) problem. In this paper, we propose a new progressive open-set recognition method with adaptive probability threshold. Both the labeled training data and the test data (objects to be classified) are put into a common data set, and the k-Nearest Neighbors (k-NNs) of each object are sought in this common set. Then, we can determine the probability of object lying in the given classes. If the majority of k-NNs of the object are from labeled training data, this object quite likely belongs to one of the given classes, and the density of the object and its neighbors is taken into account here. However, when most of k-NNs are from the unlabeled test data set, the class of object is considered very uncertain because the class of test data is unknown, and this object cannot be classified in this step. Once the objects belonging to known classes with high probability are all found, we re-calculate the probability of the other uncertain objects belonging to known classes based on the labeled training data and the objects marked with the estimated probability. Such iteration will stop when the probabilities of all the objects belonging to known classes are not changed. Then, a modified Otsu’s method is employed to adaptively seek the probability threshold for the final classification. If the probability of object belonging to known classes is smaller than this threshold, it will be assigned to the ignorant (unknown) class that is not included in training data set. The other objects will be committed to a specific class. The effectiveness of the proposed method has been validated using some experiments.