Supervised Deep Second-Order Covariance Hashing for Image Retrieval

Recently, deep hashing methods play a pivotal role in image retrieval tasks by combining advanced convolutional neural networks (CNNs) with efficient hashing. Meanwhile, second-order representations of deep convolutional activations have been established to effectively improve network performance in various computer vision applications. In this work, to obtain more compact hash codes, we propose a supervised deep second-order covariance hashing (SDSoCH) method by combining deep hashing with second-order statistic model. SDSoCH utilizes a powerful covariance pooling to model the second-order statistics of convolutional features, which is naturally integrated into the existing point-wise hashing network in an end-to-end manner. The embedded covariance pooling operation well captures the interaction of convolutional features and produces global feature representations with more discriminant capability, leading to the more informative hash codes. Extensive experiments conducted on two benchmarks demonstrate that the proposed SDSoCH outperforms its first-order counterparts and achieves superior retrieval performance.