Fast Binary Embedding of Deep Learning Image Features Using Golay-Hadamard Matrices

Convolutional neural networks (CNNs) have emerged as powerful tools for image retrieval and classification. In this paper, we study binary embedding of CNN-based image descriptors for resource-constrained devices. We propose two classes of fast computable and memory-efficient dimension reduction operators using Golay-Hadamard matrices (GHMs), which are constructed by multiplying the columns of a Hadamard matrix with a Golay sequence. Simulation results on CNN-based instance image retrieval and classification show that GHM-based operators can offer competitive performance to those of full random Gaussian matrices and Gaussian circulant matrices at much lower computational cost and storage space. This implies the potential of proposed approaches on devices with low memory, bandwidth and power restrictions.