A Locality Sensitive Low-Rank Model for Image Tag Completion

Abstract

Many visual applications have benefited from the spate of web images, but inaccurate and incomplete user tags, such as the thorn of a rose, may hinder the performance of retrieval or indexing systems that rely on this data. In this paper, we propose a new low sensitivity model for a local area to complete an image, which approximates the global nonlinear model with a set of local linear models. To effectively understand local sensitivity, a simple and effective pre-processing module was designed to learn the proper representation of data division and to introduce a global consensus system to mitigate the risk of feeding. In the meantime, the low-grade matrix factor is used as local models, where local engineering structures are maintained to obtain low-dimensional representation of both markers and samples. The large-scale empirical evaluations conducted on three data sets show the effectiveness and efficiency of the proposed method, as our way outperforms the former by a large margin.