Quantized embeddings: an efficient and universal nearest neighbor method for cloud-based image retrieval

Abstract

We propose a rate-efficient, feature-agnostic approach for encoding image features for cloudbased nearest neighbor search. We extract quantized random projections of the image features under consideration, transmit these to the cloud server, and perform matching in the space of the quantized projections. The advantage of this approach is that, once the underlying feature extraction algorithm is chosen for maximum discriminability and retrieval performance (e.g., SIFT, or eigen-features), the random projections guarantee a rate-efficient representation and fast server-based matching with negligible loss in accuracy. Using the Johnson-Lindenstrauss Lemma, we show that pair-wise distances between the underlying feature vectors are preserved in the corresponding quantized embeddings. We report experimental results of image retrieval on two image databases with different feature spaces; one using SIFT features and one using face features extracted using a variant of the Viola-Jones face recognition algorithm. For both feature spaces, quantized embeddings enable accurate image retrieval SPIE Conference on Applications of Digital Image Processing XXXVIThis work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in part without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies include the following: a notice that such copying is by permission of Mitsubishi Electric Research Laboratories, Inc.; an acknowledgment of the authors and individual contributions to the work; and all applicable portions of the copyright notice. Copying, reproduction, or republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All rights reserved. ABSTRACTWe propose a rate-efficient, feature-agnostic approach for encoding image features for cloud-based nearest neighbor search. We extract quantized random projections of the image features under consideration, transmit these to the cloud server, and perform matching in the space of the quantized projections. The advantage of this approach is that, once the underlying feature extraction algorithm is chosen for maximum discriminability and retrieval performance (e.g., SIFT, or eigen-features), the random projections guarantee a rate-efficient representation and fast server-based matching with negligible loss in accuracy. Using the Johnson-Lindenstrauss Lemma, we show that pair-wise distances between the underlying feature vectors are preserved in the corresponding quantized embeddings. We report experimental results of image retrieval on two image databases with different feature spaces; one using SIFT features and one using face features extracted using a variant of the Viola-Jones face recognition algorithm. For both feature spaces, quantized embeddings enable accurate image retrieval combined with improved bit-rate efficiency and speed of matching, when compared with the underlying featur...

Angle-preserving quantized phase embeddings

“…This, or similar, trade-offs have also been shown in the literature for quantized J-L embeddings 15,16 and universal embeddings.…”

“…Specifically, when these 2 embeddings are uniformly quantized to B bits per dimension, the embedding guarantee becomes 15,16 (…”

Quantization and Compressive Sensing

“…Embeddings are transformations that preserve the geometry of the space they operate on; reconstruction of the embedded signal is not necessarily the goal. They have been proven quite useful, for example, in signal-based retrieval applications, such as augmented reality, biometric authentication and visual search [70,21,85].…”

“…These applications require storage or transmission of the embedded signals, and, therefore, quantizer design is very important in controlling the rate used by the embedding. Indeed, significant analysis has been performed for embeddings followed by conventional scalar quantization, some of it in the context of quantized compressive sensing [54,81,82] or in the study of quantized extensions to the Johnson Lindenstrauss Lemma [55,70,85,49]. Furthermore, since reconstruction is not an objective anymore, non-contiguous quantization is more suitable, leading to very interesting quantized embedding designs and significant rate reduction [21].…”

Small Width, Low Distortions: Quantized Random Embeddings of Low-complexity Sets

“…This would lead to a fast computation of quantized mappings, with potential application in nearest-neighbor search for databases of high-dimensional signals. An open question is also the possibility to extend this work to universally-quantized embeddings [9,10,48], i.e., taking a periodic quantizer Q in (4). This could potentially lead to quasi-isometric embeddings with (exponentially) decaying distortions on vectors sets with small Gaussian width and using sub-Gaussian random matrices.…”