Abstract: Massive images are flooding into the internet in real-time in the form of streaming data, which makes the need of online image retrieval more and more urgent. To guarantee the online image retrieval performance, online Hashing algorithms are utilized to re-learn the Hash functions and re-generate the Hash codes of the new and old samples in real time. As time went by, the amount of old dataset is very large, and the time complexity of re-generating Hash codes become unacceptable. To avoid the above problems, we propose a novel asymmetric deep online Hashing (ADOH) which trains a deep online Hashing network in an asymmetric manner. To improve the online retrieval efficiency, ADOH only generates the Hash code of the new samples and do not update the old samples’ Hash codes. During the training process, ADOH minimizes the difference between the Hash codes’ inner product and the similarity matrix, which can preserve the pair-wise semantic similarity relationship. Moreover, ADOH establishes the classification loss and the label embedding model to learn the samples’ semantic information, which makes the generated Hash codes more semantically discriminative. We conduct the approximate nearest neighbor retrieval comparative experiments on three widely used datasets including the cifar-10 dataset, the mnist dataset and the Places205 dataset. The results show that the online approximate nearest neighbor retrieval performance of ADOH outperforms the other 8 existing online Hashing methods.