Clustering Based Uncorrelated Balanced Class-specific Learning for Image Classification
Fei Wu · Journal of Information and Computational Science · 2015
Class-specific learning is an effect feature learning technique. Yet it tends to recast traditional multi-class feature learning problem into several binary class problems, and suffers from the class-imbalance problem inevitably, where the minority class is the specific class and the majority class consists of all the other classes. Although the Class-balanced Discrimination (CBD) and Orthogonal CBD (OCBD) methods attempt to address this problem, there still exists much room for improvement. In this paper, we propose a novel class-specific learning approach named Clustering based Uncorrelated Balanced Class-specific Learning (CUBCL). For a specific class, we select a reduced counterpart class whose data are nearest to the data of specific class. We employ the K-means clustering technique to divide the reduced counterpart class data into appropriate number of subsets, which can bring favorable clustering performance for each subset. Each subset is then combined with the specific class to form a balanced group for discriminant feature learning. Furthermore, we design the locally statistical uncorrelated constraints for class-specific learning, so as to remove the local correlations of discriminant features obtained from multiple groups. Experiments on two public datasets demonstrate the effectiveness of the proposed approach.