Aiming at the disturbance of training samples with large aspect gap in SAR target recognition, a kernel function transformation collaborative algorithm based on adaptive atom selection is proposed. This method improves the representation dictionary in the traditional collaborative representation, and gets the adaptive dictionary which is more adaptable to the current test samples and can reduce the influence of the unrelated atoms to the system. The experiments of SAR target recognition based on MSTAR datasets were carried out. The experimental results show that the kernel collaborative representation based on adaptive atom selection is more effective than the kernel collaborative representation model based on all training sample dictionary, which reduces the harmful effect of the interference atoms and further improves the reliability and robustness of the system.