Survey of Hashing Techniques for Compact Bit Representations of Images

Binary encoding schemes that preserve similarity in large collections of images are required for faster retrieval and effective storage. There have been many recent hashing techniques that produce semantic binary representations. This paper presents a survey of such hashing techniques that allows faster nearest neighbor search in hamming space. Specifically, approaches that use locality-sensitive hashing family functions, spectral relaxation or k-means like iterative schemes, and those that minimize the quantization/reconstruction loss between the original metric and the hamming space are presented. We evaluate the binary codes obtained using these methods with 4 protocols: mean average precision, precision-recall and precision and radius with hamming radius less than a certain amount on 4 small datasets1: labelme, notredame, peekaboom and mnist and 3 bigger datasets: Labelme22k, Cifar-10 and Cifar-100 to measure the performance. We show that Spherical Hashing works well on big datasets and MLH, KSH, BRE perform better on small datasets. We also evaluate the methods by changing the definition of ground truth neighbors and show that the performance changes in this scenario.