Kernel Discriminant Analysis Using Triangular Kernel for Semantic Scene Classification

Semantic scene classification is a challenging research problem that aims to categorise images into semantic classes such as beaches, sunsets or mountains. This problem can be formulated as multi-labeled classification problem where an image can belong to more than one conceptual class such as sunsets and beaches at the same time. Recently, kernel discriminant analysis combined with spectral regression (SR-KDA) has been successfully used for face, text and spoken letter recognition. But SR-KDA method works only with positive definite symmetric matrices. In this paper, we have modified this method to support both definite and indefinite symmetric matrices. The main idea is to use LDL^T decomposition instead of Cholesky decomposition. The modified SR-KDA is applied to scene database involving 6 concepts. We validate the advocated approach and demonstrate that it yields significant performance gains when conditionally positive definite triangular kernel is used instead of positive definite symmetric kernels such as linear, polynomial or RBF. The results also indicate performance gains when compared with the state-of-the art multi-label methods for semantic scene classification.