Research On Salient Region Detection Based On The Learning Of Feature Distribution

Image salient region detection is to extract the salient object from a visual scene by imitating the human visual attention mechanism,so as to let machine have visual initiative and selective like human beings.It has many applications such as image segment,image retrieval and image/video compression.Traditional salient region detection methods integrate features by heuristic function,which need empirical knowledge.Machine learning methods can automatically optimize models by learning the data,science learning needs sufficient data,we can integrate more features.In this paper,we explain the basic concepts of salient region detection,and make an overview of some existing methods,then we expound the principle of supervised learning.In order to improve the ability of feature integration and the automation of detecting,we propose two kinds of salient region detection methods based on the learning of feature distribution.They share the way of combining supervised learning model with semi-supervised learning model:(1)Supervised learning model,which is used to generate the coarse salient map.Both two kinds of methods use the Gradient Boosted Decision Tree,which can handle feature vector with very high dimension,and choose features according to their classification ability.(2)Semi-supervised learning model,which is used to optimize the coarse salient map.The first kind of method uses Label Propagation combining with Gaussian Mixture Model,after Label Propagation process regions with low salient confidence,Gaussian Mixture Model classifies regions by calculating probabilities that they belong to each categories.The second kind of method improves Label Propagation by combining adjacent similarity with nearest neighbor similarity,which allows us to abandon Gaussian Mixture Model,and thus ameliorates the too-detailed disadvantage of the first kind of method.We compare the two proposed methods with six other salient region detection methods.Experiments show that our methods is better than the six other methods,and the second proposed method is better than the first one.At the end of this paper,we summarized our work and made a prospect.