Image salient region detection based on the learning of feature distribution
Xian Jun Shi, Jin Wu, Lei Zhu · 2017
Traditional salient region detection methods mostly incorporate heuristic functions, and the result always depends largely on the parameters of model. On the other hand, statistical learning based methods can learn the model automatically. In this paper, we propose a learning approach to detect the salient region of an image from coarse to fine. Firstly, we segment images into multi-scale superpixels, and train a gradient boosted decision tree to predict the coarse salient map. Secondly, we employ label propagation method to process unlabeled superpixels. Finally, a Gaussian Mixture Model is built to generate the final salient map. The AUC score of our method on MSRA5000 is 0.8774, which shows that our method can detect salient region accurately. In addition, few parameters are required in our system, which makes our method more practical in real applications.