The main idea of Matching Pursuit (MP) is to get a local optimal solution by iteration, so as to gradually approach the original signal. To cope with the intersection of different atom sets, which may affect the classification performance of conventional MP methods, a new matching pursuit algorithm is proposed, which is suitable for supervised classification. The criterion for atoms selection consists of two parts. On one hand, by using the same atom set within the class, the intra-class structure of the similar signals is obtained for class-representation; on the other hand, by selecting the atom sets independently for every class, the discrimination ability for different classes could be further strengthened. The analysis on a toy example indicates that this scheme reduces the common factors between different classes and highlights the discrimination between signals, which may boost the performance of signal classification. Finally, the experiments on benchmark image databases and the measured radar emitter signals verify that the proposed algorithm achieves better robustness against noise and occlusion, compared with the convention MP-related methods.