Complicated scene retrieval using block voting mechanism and weak feature selection based on bag-of-features

摘要

In this paper, we focus on the problem of complicated scene retrieval and give two proposals to improve accuracy of the recent image search system based on bag-of-features: block voting mechanism and weak feature selection. Both the methods aim to reduce effects of incorrect matching between descriptors. Block voting mechanism separates query and database images into blocks when computing image matching scores. It can be integrated into inverted file for an efficient and compact indexing structure. Weak feature selection provides a simple approach to select good feature points for matching. Experiments performed on a dataset with complicated scene and various transformations including viewpoint and illumination changes show an about 20 percent improvement rather than baseline bag-of-features due to my proposals.