Partial Random Spherical Hashing for Large-scale Image Retrieval

Hashing-based searching approaches have been widely employed in large-scale image retrieval tasks. However, most hashing schemes are developed based on hyperplane projection, which may not effectively capture the spatially coherent data structure. More importantly, the existing approaches often compromise between learning efficiency and retrieval accuracy, and can thus barely satisfy the real-time requirements. In this paper, we propose a novel hashing method, which is referred to as Partial Random Spherical Hashing (PRSH), for large-scale image retrieval. First, the images are encoded into a lower hamming space via some randomly generated hyperspheres. Then, a fast learning scheme is adopted to adjust the codes and make them have universal approximation capability to the original image features. The co-play between randomness and learned parameters results in a both efficient and effective learning scheme for constructing hash functions. Experiments on two large public image datasets have shown that our PRSH method outperforms state-of-the-arts in terms of both learning efficiency and retrieval accuracy.