Hashing has been widely used in large-scale vision problems thanks to its efficiency in both storage and speed. For fast cross-modal retrieval task, cross-modal hashing (CMH) has received increasing attention recently with its ability to improve quality of hash coding by exploiting the semantic correlation across different modalities. Most traditional CMH methods focus on designing a good hash function to use supervised information appropriately, but the performance are limited by hand-crafted features. Some deep learning based CMH methods focus on learning good features by using deep network, however, directly quantizing the feature may result in large loss for hashing. In this paper, we propose a novel end-to-end deep cross-modal hashing framework, integrating feature and hash-code learning into the same network. We keep the relationship of features between modalities. For hash process, we design a novel net structure and loss for hash learning as well as reconstruct the hash codes to features to improve the quality of codes. Experiments on standard databases for cross-modal retrieval show the proposed methods yields substantial boosts over latest state-of-the-art hashing methods.