Research On Visual Saliency Model And Eyetracking Analysis In Target Search

When faced with the complicated environment, our eyes can always fixate at the salient target rapidly. With the development of biology, psychology and computer science in recent years, more and more researchers contribute to the computational model of visual saliency system. This article will study the mechanism of our visual saliency system from the view of model theory and experiment.Firstly, we propose a bottom-up model to simulate the process of our visual system. We extract high-order features from current image based on the sparse coding method. Then we propose a sparse center-surround model to depict the local saliency based on the global sparse code. The saliency map is combined with the high-order and low-order feature map. The result shows that our model can achieve a better performance in saliency prediction than the classical models.However, the prior knowledge of target will have an effect on our attention result. So we propose a top-down model based on hidden semantic information to investigate this mechanism. We extend the 1D Hidden Markov Model to 2D-model and propose a 2D-viterbi algorithm based on 1D-viterbi algorithm to adapt this model. The result shows that our semantic model can predict the salient region and the shape of target, and what’s more, our model is robust to the imperfect targets’ shape.In order to test our models, we take some experiments with them in target search. We build a framework based on the bottom-up model to detect salient objects, and our model achieves a good result. Then we take two eye track experiments with EyelinkⅡ device to study the relation between our models and subjects. The first experiment shows that our bottom-up model is similar to subjects in searching task, such as the difficulty of image understanding. Moreover, our model can achieve a better performance than subjects. On the other hand, we find the fixation data of subjects implies much information. So we take another eye track experiment and use the top-down models to learn from the eye data, the result shows that our model can fetch the targets, which proves the significance of the eye track data.The validity of our two proposed models indicates the saliency of local region based on global context is effective and the experiment denotes that eye track data can be exploited to make sense of our visual system.