Content-Based Image Retrieval by Query Adaptive Search using Hash Codes

Abstract: Visual similarity has been a lively topic of analysis in Scalable image search. Progressive solutions usually use hashing ways to plant high-dimensional image options into playing area, wherever search may be performed in period supported playing distance of compact hash codes. Not like ancient metrics (e.g., Euclidean) that provide continuous distances, the playing distances area unit separate number values. As a consequence, there area unit usually an outsized range of pictures sharing equal playing distances to a question that mostly hurts search results wherever fine-grained ranking is vital. This paper introduces associate approach that allows query-adaptive ranking of the came pictures with equal playing distances to the queries. This can be achieved by first off offline learning bitwise weights of the hash codes for a various set of predefined linguistics thought categories. We have a tendency to formulate the burden education method as a quadratic programming drawback that minimizes intra-class distance whereas conserving inter-class relationship captured by original raw image options. Query-adaptive weights area unit then computed on-line by evaluating the proximity between a question and also the linguistics thought categories.

Keywords: Hash codes, Query-adaptive, Hamming distance.