Projection selection hashing

Hashing methods can perform large scale image retrieval tasks efficiently. Random projection hashing methods, such as the SKLSH, have excellent performances in long code length situation. Owing to the randomness of the projections of random projection hashing methods, projections can be `good´ or `bad´. In this work, we propose the Projection Selection Hashing (PSH) to evaluate qualities of projections and select projections for the binary coding to yield a better random projection in comparison to the SKLSH. The PSH is a two-phase algorithm. In the first phase, real-valued vectors describing images are mapped to binary codes by using the random Fourier features projection method. In the second phase, projections with high quality by are selected. The proposed projection selection algorithm is motivated by the feature selection algorithm Relief. We extend it for unsupervised problems. Then, we propose a new update rule to estimate the quality of projections.