A dynamic dictionary learning method was proposed for automatic target recognition in Synthetic ApertureRadar (SAR) images. The new method decreased the size of a dictionary by erasing the useless or high-correlation atoms in the dictionary under a deleting cost. It added an atom by decomposing the residue data matrix under an adding cost. The performance of signal representation of a dictionary could be improved in the dynamic learning procedure. Experiments based on MSTAR database show that the proposed method can converge to a proper-size dictionary no matter how large the original size is. The proposed method was demonstrated to have good and robust performance of SAR target recognition in the trials.