Kernel Homotopy based sparse representation for object classification

The l₁ minimization problem (Lasso) is a basic and critical problem in sparse representation and its applications. Among the solutions, Homotopy is an efficient and effective algorithm. In this paper, we propose a novel kernel algorithm based on Homotopy (KHomotopy) to solve the Lasso problem in kernel space. Then we integrate it in the well known Sparse Representation based Classification (SRC) framework. The proposed method is applied to the object classification problem, and compared with other kernel SRC methods and kernel SVM. Experiments on the CalTech101 and the Flower 17 databases show that KHomotopy has the best overall performance in accuracy and speed, which outperforms both linear SRC and KSVM, and is better than or comparable to two existing kernel SRC algorithms.