RSF: A Novel Saliency Fusion Framework for Image Saliency Detection

Image saliency detection has become a hot topic in computer vision tasks. A large number of saliency models have been developed in recent years. Considering that these models use different prior knowledge, features, and theoretical methods, they have their own strengths and weaknesses in different images. Based on this observation, some works aim to fuse multiple weak saliency models based on different fusion strategies (Each saliency model in fusion framework is called a weak saliency model). However, these fusion methods lose effectiveness when image content is very complex, because they ignore the fact that various regions in an image have different characteristics in saliency fusion. Different with them, we propose a novel Region-level Saliency Fusion framework (RSF) by exploring the relationship between weak saliency models and image regions. For the input image and J weak saliency models, we firstly segment input image into N regions. Then, our goal is to learn to infer the reliability of using each weak saliency model to predict each image region saliency value. This way, we can select more reliable weak saliency models for each region to predict its saliency value. Finally, we use smoothness prior to further smooth the saliency map obtained by RSF. Experimental results on three datasets demonstrate the superiority of the proposed method than other state-of-the-art methods.