Second-Order Pooling Deep Hashing For Image Retrieval

Recently, due to advantages of high computational efficiency, small storage cost as well as high discriminability, deep hash methods have been widely studied in a number of large-scale visual applications. To achieve more compact hash representation, we propose a novel supervised deep hash method for image retrieval task in this work, which successfully embeds second-order pooling operation into existing deep hash model in an end-to-end manner, namely second-order pooling deep hashing (SoPDH). Our SoPDH mainly consists of four parts, i.e., a basic deep feature extraction module, a second-order pooling operation based on matrix decomposition, a hash encoding module and a semantic classification layer. The embedded second-order pooling operation not only guarantees the local prominence of deep features, but also introduces the global statistic feature information, which could lead to a more robust hash coding result. We extensively evaluate the proposed SoPDH on two commonly-used datasets, and experimental results demonstrate its effectiveness for image retrieval task.