Research Of Visual Attention Model Based On Graph Spectra

Visual attention model,which is based on studies and hypotheses on neural psychology, epistemology and anatomy ,uses math model to simulate human visual system. Visual attention model detects the regions of interest by detecting prominent points in an image,for it takes the regions of interest which contain potential target as the muster of points that have prominent features. It has become a hot topic in the digital image processing field. In image processing tasks such as target detection, surveillance and image retrievel,it would be more convenitent to focus on regions of interest detected by visual attention model than focus on the whole image.Graph theory based image processing is newly developing technique in recent years.It is developed greatly in the image segmentation.The image is mapped into a weighted undirected graph and the pixels are considered as vertexes and the similarity between the visual properties(e.g.graylevel intensity,color or texture)at each pair of neighboring pixels is assigned as the respective edge weight.Therefore the image processing can be obtained by computing the graph with some arithmetices.This thesis studies on bottom-up visual attention model. The shortage of both Itti's model and Harel's model is investigated and a new model is proposed, the visual attention model based on computing degree. Main works and content of this paper include:1. This thesis introduces visual attention mechanism and progresses that have been made in the field. Physiology and psychology studies related to visual attention and both Itti's model and Harel's model are also introduced.2. This thesis introduces detailedly visual attention model based on computing degree and investigates the computal process of this model, including strategies of the detection of primary visual features and combination of conspicuity maps.3. To reduce computational complexity, this model needs three floors of the pyramid structure data to detect primary visual features, while Itti's model needs nine floors and Harel's model needs four floors.4. To get saliency map, feature maps is computed by the arithmetic of computing degree which is sample on principle. Its realization is easy and its computal effect is better than others.It shows in our experimentation that the improved model proposed by this thesis can effectively detect regions of interest and compared to Itti model, the new model can improve the detection of regions of interest. Meanwhile the computation is not too complex.