Hierarchy Features Based Visual Attention Research

When looking at an image, people may focus on some areas and ignore certain areas without consciousness. This selective mechanism in visual perception process is the result of visual attention. In computer vision research, computer can get the ability to automatically obtain the human visual region of interest in a complex environment by modeling of visual attention mechanism.By simulating the human visual perception system, researchers had proposed a feature integration based visual attention computational framework. A variety of visual attention model derived from this framework. This article analysis Itti and Judd visual attention models which had significant influence in this research area. Itti model generated saliency map as prediction of the degree of concern in an image by integrating a variety of low-level features. This method ignored the fact that visual attention process is inevitably influenced by knowledge, tasks, preferences, and other factors. Judd model added high-level semantic features as way of introducing knowledge. Despite the good results, the design and calculation of heuristic features is complex and model's scalability is not strong.This paper focuses on two issues based on the existing model of visual attention:(1) How to obtain the features of visual attention through learning.(2) How to integrate hierarchy features under the framework of feature integration. Firstly, this paper gets pixel-level, object-level, semantic-level features by training convolution neural network. Then, based on acquired learning features, a new visual attention model which integrate hierarchy features is proposed. This model emphasis on high-level features integration by using object attribute information, which effectively compensate for the lack of existing models in the introduction of priori knowledge. Finally, for the proposed visual attention model, an attention focus shift method is designed which is guided by hierarchy knowledge. Experiments show that the new model takes full advantage of prior knowledge and good results are obtained by testing on multiple datasets.