Class-imbalance learning based discriminant analysis

Feature extraction is an important research topic in the field of pattern recognition. The class-specific idea tends to recast a traditional multi-class feature extraction and recognition task into several binary class problems, and therefore inevitably class imbalance problem, where the minority class is the specific class, and the majority class consists of all the other classes. However, discriminative information from binary class problems is usually limited, and imbalanced data may have negative effect on the recognition performance. For solving these problems, in this paper, we propose two novel approaches to learn discriminant features from imbalanced data, named class-balanced discrimination (CBD) and orthogonal CBD (OCBD). For a specific class, we select a reduced counterpart class whose data are nearest to the data of specific class, and further divide them into smaller subsets, each of which has the same size as the specific class, to achieve balance. Then, each subset is combined with the minority class, and linear discriminant analysis (LDA) is performed on them to extract discriminative vectors. To further remove redundant information, we impose orthogonal constraint on the extracted discriminant vectors among correlated classes. Experimental results on three public image databases demonstrate that the proposed approaches outperform several related image feature extraction and recognition methods.