Class-noise Tolerant Classification Based on a Probabilistic Noise Model

Class noise usually means the erroneous labeling of the training examples. In pattern recognition problems, class noise occurs frequently and deteriorates the classifier derived from the noisy dataset. For instance, in some adaptive image segmentation system, the class labels of the training pixels are assigned automatically and sometimes contain errors. Since image segmentation plays a crucial role as a preliminary step for high level image processing, the class noise problem needs to be addressed. This paper presents several possible solutions to this problem based on a probabilistic noise model (LSA). These solutions include the Clustering-based Probabilistic Algorithm (CPA), the Probabilistic Fisher method (PF), and the Probabilistic Kernel Fisher method (PKF). The proposed algorithms enable standard classifiers to tolerate class noise, and extend the earlier work [1] in several ways. The experimental results show that the proposed approaches improve standard classifiers in noisy datasets.