Scalable Image Retrieval by Sparse Product Quantization

TL;DR

This paper introduces Sparse Product Quantization (SPQ), a novel method for high-dimensional image feature encoding that reduces quantization errors and improves approximate nearest neighbor search performance in large-scale image retrieval.

Contribution

The paper proposes a sparse encoding approach for product quantization, significantly reducing quantization errors and enhancing retrieval accuracy over traditional methods.

Findings

Achieved state-of-the-art results on four public image datasets.

Demonstrated effective compression and encoding of high-dimensional features.

Validated improved retrieval performance in content-based image retrieval.

Abstract

Fast Approximate Nearest Neighbor (ANN) search technique for high-dimensional feature indexing and retrieval is the crux of large-scale image retrieval. A recent promising technique is Product Quantization, which attempts to index high-dimensional image features by decomposing the feature space into a Cartesian product of low dimensional subspaces and quantizing each of them separately. Despite the promising results reported, their quantization approach follows the typical hard assignment of traditional quantization methods, which may result in large quantization errors and thus inferior search performance. Unlike the existing approaches, in this paper, we propose a novel approach called Sparse Product Quantization (SPQ) to encoding the high-dimensional feature vectors into sparse representation. We optimize the sparse representations of the feature vectors by minimizing their quantization errors, making the resulting representation is essentially close to the original data in practice. Experiments show that the proposed SPQ technique is not only able to compress data, but also an effective encoding technique. We obtain state-of-the-art results for ANN search on four public image datasets and the promising results of content-based image retrieval further validate the efficacy of our proposed method.