Filtering out background features from BoF representation by generating fuzzy signatures
Qiang Qiu, Qixin Cao, Masaru Adachi · 2014
Bag of Features (BoF) approach has gained its popularity in the past decade due to its simplicity and outstanding performance in computer vision tasks. However, the lack of spatial information makes the BoF method sensitive to background noise features in real-world object recognition tasks. This paper presents a method for removing background noise features from the BoF representation of images by generating fuzzy signatures. This technique treats each visual word as a fuzzy set, and defines a membership function to wipe off background features in testing images. The experimental results show that fuzzy signature can remove some background features and improve the performance of BoF method in real-world object recognition tasks.