Research Of Image Annotation By Sparse Regression Model

Image annotation is the basis of large-scale image retrieval technology. It is one of the hottest research topics in Multimedia. Image annotation aims at assigning several key-words to images to describe the visual information included in them. In common application, one image can be annotated with several different key-words, so single-label image annotation can be extended to a multi-label problem, both of which are widely studied and used.First, this paper introduces methods commonly used for image annotation and multi-label image annotation. This paper also introduces the way to regularize regression model with sparse constraints.This paper proposes two models for image annotation and imposes l1 sparse constraint term in to on regression models which make the models more interpretable. These two annotation models are sparse logistic regression model and two-layer sparse logistic regression model.The sparse logistic regression model encodes the relationship between visual features in images and labels with logistic function. Given the visual features of image, the sparse logistic regression model computes the conditional probability of corresponding image label. Generally, considering that every label is only correlated with a small number of visual words, the model introduces in l1 sparse constraint term to regularize the parameters of the model and make it more interpretable.The two-layer sparse logistic regression model has two layers. The first layer separately learns the relationship between visual features and image labels as well as the relationship between different labels. The second layer combines the two different kinds of relationships together and employs both of them in multi-label image annotation. Similarly as that every label correlates with only a small number of visual words, it also correlates with only a small part of the whole set of image labels. Therefore, the two-layer sparse logistic regression model separately introduces in two l1 sparse constraint terms to regularize the parameters of the regression model and make the model more interpretable.