Traditional hand-crafted features for representing local image patches are evolving into current data-driven and learning-based image feature, but learning a robust and discriminative descriptor which is capable of controlling various patch-level computer vision tasks is still an open problem. In this work, we propose a novel deep convolutional neural network (CNN) to learn local feature descriptors. We utilize the quadruplets with positive and negative training samples, together with a constraint to restrict the intra-class variance, to learn good discriminative CNN representations. Compared with previous works, our model reduces the overlap in feature space between corresponding and non-corresponding patch pairs, and mitigates margin varying problem caused by commonly used triplet loss. We demonstrate that our method achieves better embedding result than some latest works, like PN-Net and TN-TG, on benchmark dataset.