Kernel correlation filter for vehicle detection and recognition in SAR images

SAR target detection and recognition is sensitive to target's azimuth. To solve the problem, based on correlation theory and kernel feature analysis, a kernel correlation filter which is strongly robust to target's azimuth distortion is proposed. The novel filter exploits eigenvectors to reduce the dependence of the training set and extends linear combination of eigenvectors nonlinearly to improve the classification. Moreover, to keep the computation tractable in high dimensional space, the kernel function is employed. Comparative tests using MSTAR database demonstrate the kernel correlation filter performs high detection probability with low false alarm probability and implements target detection and recognition accurately without templates and target poses estimation.