Fast and Randomized Multiple Kernel Discriminant Analysis for Bird Recognition

In order to improve the bird strike avoidance management at airports, and to realize the linkage of bird detection radar and a variety of bird repellent equipment, intelligent bird repellent decision methods have attracted widespread attention. In this paper, we are committed to exploring bird target recognition approaches, so as to prepare for the subsequent bird repellent decision methods. Kernel methods are well known to be effective in dealing with nonlinear machine learning problems. However, the performance of kernel methods heavily relies on the choice of kernel parameters. To mitigate this problem, we add the ideas of multiple kernels and randomized techniques to kernel discriminant analysis (KDA), and propose a randomized and fast multiple kernel discriminant analysis (RM-KDA) method with applications to bird recognition. The main contributions of our work are two-fold. First, we propose an innovative multiple kernel discriminant analysis (M-KDA) framework to effectively solve recognition tasks with nonlinear data structures. Second, we develop a fast and efficient randomized algorithm to solve the M-KDA model without explicitly forming and storing the base kernel matrices and the ensemble kernel matrix in advance, which greatly reduces the computation and storage requirements. The extensive experiments on benchmark recognition databases demonstrate the effectiveness of the proposed algorithm.