Similarity Measurement Based on Non-linear Hash Coding

Abstract

We propose an algorithm named non-linear deep hash (NLDH) to encode the object in the image into a series of compact binary codes through a structure called a non-linear hash coding module. Based on this, then, we propose a retrieval algorithm for similar images based on these binary codes in the image. This algorithm adopts the strategy from coarse-to-fine. Then the image level similarity calculation is carried out to complete the search of the most similar images. Finally, experiments were carried out on the Oxford buildings dataset and ancient painting image dataset in this paper. The experimental results show that our proposed algorithm has a higher retrieval ability than the ordinary deep hash method, and the retrieval accuracy and recall rate are greatly improved.