To efficiently retrieve in multimodal data, it is essential to reduce the proportion of irrelevant documents. The image data were projected to the Hamming space by using the localitysensitive hashing algorithm, the text data were mapped on the hashing function of Hamming space by employing the neural network learning, and then a novel crossmedia retrieval approach was proposed to reduce the proportion of irrelevant documents. The experiment shows that the proportion of the relevant documents can be much improved in the proposed method. Assessments on the two public datasets also demonstrate the efficacy and the accuracy of the proposed retrieval method when compared to the baselines.