Large-scale image retrieval based on boosting iterative quantization hashing with query-adaptive reranking

Image hashing based Approximate Nearest Neighbor (ANN) searching has drawn more and more attention in large-scale image dataset applications. It is still challenging to learn hashing codes to achieve good search performance. In this paper, we propose an image retrieval method based on boosting iterative quantization hashing method with query-adaptive reranking. Firstly, in boosting iterative quantization hashing embedding, we adopt boosting-based method to generate inputs to learn hashing functions. Then we optimize the hashing functions with a loss function by considering the relationship between samples. Once the hashing codes are generated, Query-Adaptive Reranking (QAR) method is proposed to learn bit-level weights for each category and query-adaptive weights for each hashing bit. In this way, the discrete Hamming distance value can be continuous, and many irrelevant returned images can be sorted to the back. We conduct experiments on three public datasets, and comparison results with six state-of-the-art methods to illustrate the effectiveness of the proposed method.