Due to limited features extracted from compression domain, the conventional Euclidean distance based retrieval performance in compressed-domain is not satisfactory. The Distance Metric Learning(DML) is introduced to compressed-domain images retrieval and a DML based Discrete Cosine Transform(DCT) domain retrieval for Joint Photographic Experts Group(JPEG) images is developed. Firstly, we propose a more effective DCT domain features extraction method, and then the DML is applied to train a more efficient metric matrix for retrieval. Retrieval experiment on Corel5000 images database demonstrates that the approach proposed can effectively improve the retrieval accuracy.