Abstract: In order to improve the retrieval accuracy and reduce the computational cost, a cross modal hash retrieval method based on subspace relation learning is proposed. By optimizing the distance between the hash code and the relation information, the class label was transformed into the relation information of the subspace, thus preserving the modal relation, discrete constraint and nonlinear structure. A symmetric framework was designed to generate a unified binary code retrieval database, and a discrete optimization Hash algorithm was proposed to solve the problem that the objective function did not relax the discrete constraints, which effectively improved the training efficiency. Two cross modal retrieval experiments show that the proposed method has higher retrieval accuracy and less computation time.