Deep Spherical Quantization for Image Search

Abstract

Hashing methods, which encode high-dimensional images with compact discrete codes, have been widely applied to enhance large-scale image retrieval. In this paper, we put forward Deep Spherical Quantization (DSQ), a novel method to make deep convolutional neural networks generate supervised and compact binary codes for efficient image search. Our approach simultaneously learns a mapping that transforms the input images into a low-dimensional discriminative space, and quantizes the transformed data points using multi-codebook quantization. To eliminate the negative effect of norm variance on codebook learning, we force the network to L 2 normalize the extracted features and then quantize the resulting vectors using a new supervised quantization technique specifically designed for points lying on a unit hypersphere. Furthermore, we introduce an easyto-implement extension of our quantization technique that enforces sparsity on the codebooks. Extensive experiments demonstrate that DSQ and its sparse variant can generate semantically separable compact binary codes outperforming many state-of-the-art image retrieval methods on three benchmarks.

Webly Supervised Deep Attentive Quantization

“…Besides, we use the pre-trained Word2Vec [17] model to embed each tag as a 300-dimensional vector. Following [10,11,28], we adopt K = 256 codewords for each codebook, thus the binary quantization code for each image of all M codebooks requires B = M log 2 K = 8M bits (i.e., M bytes).…”

Lossy Compression of General Random Variables

“…) • For the class of m-rectifiable random variables [40, Definition 11], Proposition 3.1 recovers [40, Theorem 55]. In this case, h µ (X) is the m-dimensional entropy [40, Definition18] of the m-rectifiable random variable X.…”

“…In this section, we particularize our results on R-D theory and quantization to random variables taking values in compact manifolds, specifically hyperspheres and Grassmannians. Hyperspheres are prevalent in many areas of data science including spherical quantization [58,48,18], hypersphere learning [34, Section 4], and directional statistics [47]. Grassmannians find application in, e.g., code design [65,30,13,51], computer vision [59], deep neural network theory [32], and the completion of low-rank matrices [7].…”

Challenges, Opportunities, and Future Prospects

“…Consequently, it is highly suggested to develop a new federated learning scheme to take into consideration the trade-off between utilization of communication resources and convergence rate throughout the training phase. Numerous quantization methods might be applied for federated learning across IoT networks including hyper-sphere quantization [78], low precision quantizer [79,80], and universal vector quantization [81].…”