Research And Application Of Visual Attention Model

Visual attention is the ability that human focuses attention on a few important areas, and then utilizes limited computational resource to process them prior, when facing a complex scene. However, many image processing algorithms waste much time and computational resource to deal with complex scene at the moment. So it's important to model the visual attention and introduce it into the procedure of image processing.First, this dissertation introduces the physiological structures of visual attention. Each part of vision is explained along the visual path. Then a few important'top-down'visual attention models are introduced, including Itti's model which models the physiological structures strictly and is used most widely, Spectral Residual algorithm which is most efficient, PQFT algorithm which is expanded from Spectral Residual, and the global contrast algorithm which is the latest and best performance.However, all these algorithms have drawback. For example, Itti's model is inefficient; Spectral Residual algorithm doesn't utilize color information; Spectral Residual algorithm and PQFT algorithm tend to highlight the boarder of object; the global contrast algorithm can't process the image with complex background.At last, this dissertation proposes a new salient region detection algorithm. The goals of the algorithm are:1. Uniformly highlight the whole salient regions instead of edges or boarders.2. Be able to detect salient object from background with complex texture.3. Be efficient, so as to meet the demand of large-scale image processing, retrieval and classification. After analysis, this dissertation proposes the frequency and spatial domain analysis based salient region detection method. We find out that the background which attract less attention usually appear periodically, and lead to sharp spikes in the amplitude spectrum. We remove the spikes in the amplitude spectrum by median filter and thereby suppress the periodical background. How to choose the window size of median filter is a problem. First we filter the amplitude spectrum with different window size, and transform all these results into spatial domain. Then we utilize spatial standard deviation to select the candidate. At last, we define and utilize a contrast function to choose the best saliency map. Our algorithm utilizes the intensity and color information, and utilizes global contrast algorithm to improve the result.The frequency domain analysis aims at detecting salient object from complex background, and it's very efficient. Meanwhile, the spatial domain analysis covers the shortage of frequency domain analysis which highlights the boarder of object, and highlights the whole object uniformly. So the goals are achieved.We compare our algorithm and 10 other state-of-the-art algorithms on MSRA salient object database and the largest accurately labeled salient object dataset. The results show that our algorithm outperforms other methods on precision rate, recall rate, F-measure, and the area under ROC.