A Compression Hashing Scheme for Large-scale Face Retrieval

Hashing method has the intrinsic problem that a long binary code yields better precision but requires a larger storage cost. Most of existing hashing methods aim to find an optimal code length to trade off the precision and storage. However, in reality, the scale of the face images is enormous and thus the storage burden is unimaginative heavy. We propose to apply a similarity-preserving compression scheme to existing unsupervised hashing methods, so as to reduce storage burden while maintaining a high precision. We employ two different lengths of code, including a long code with original length and a short code with length after m-time compression. The hash code for the query face preserves the original code length while the hash code for stored image is compressed with a ratio m to reduce storage cost. When performing face retrieval, the compressed hash code for the stored face is m-time repeatedly concentrated, in order to be compared with the long hash code for the query based on Hamming distance. Experimental results on large-scale retrieval demonstrate that the proposed compression scheme can be efficiently applied in existing methods and achieves both a high precision and a small storage space.