Online variable topology type photovoltaic grid-connected inverter
Learning to rank using user clicks and visual features for image retrieval
1. LEARNING TO RANK USING USER CLICKS AND VISUAL FEATURES FOR IMAGE
RETRIEVAL
ABSTRACT
The inconsistency between textual features andvisual contents can cause poor image
search results. To solvethis problem, click features, which are more reliable than
textualinformation in justifying the relevance between a query andclicked images, are adopted in
image ranking model. However, theexisting ranking model cannot integrate visual features,
which areefficient in refining the click-based search results. In this paper,we propose a novel
ranking model based on the learning to rankframework. Visual features and click features are
simultaneouslyutilized to obtain the ranking model. Specifically, the proposedapproach is based
on large margin structured output learningand the visual consistency is integrated with the click
featuresthrough a hypergraphregularizer term. In accordance with thefast alternating linearization
method, we design a novel algorithmto optimize the objective function. This algorithm
alternately minimizestwo different approximations of the original objective functionby keeping
one function unchanged and linear zing the other.We conduct experiments on a large-scale
dataset collected fromthe Microsoft Bing image search engine, and the results demonstratethat
the proposed learning to rank models based on visualfeatures and user clicks outperforms state-
of-the-art algorithms.