Image Retrieval via Balanced and Maximum Variance Deep Hashing

Hashing is a typical approximate nearest neighbor search approach for large-scale data sets because of its low storage space and high computational ability. The higher the variance on each projected dimension is, the more information the binary codes express. However, most existing hashing methods have neglected the variance on the projected dimensions. In this paper, a novel hashing method called balanced and maximum variance deep hashing (BMDH) is proposed to simultaneously learn the feature representation and hash functions. In this work, pairwise labels are used as the supervised information for the training images, which are fed into a convolutional neural network (CNN) architecture to obtain rich semantic features. To acquire effective and discriminative hash codes from the extracted features, an objective function with three restrictions is elaborately designed: (1) similarity-preserving mapping, (2) maximum variance on all projected dimensions, (3) balanced variance on each projected dimension. The competitive performance is acquired using the simple back-propagation algorithm with stochastic gradient descent (SGD) method despite the sophisticated objective function. Extensive experiments on two benchmarks CIFAR-10 and NUS-WIDE validate the superiority of the proposed method over the state-of-the-art methods.