Research On Applications Based On Visual Saliency

Nowadays, the requirement for the applications based on visual data, such as images and videos, is growing rapidly because of the development of network and explosive growth of visual data. We are more inclined to get information from the images and videos. We dream that the mechines can obtain useful information from these data quickly and accurately, like a man, and research on Visual Saliency is carried out to achieve the goal.Visual Saliency is a visual congition mechanism to direct visual attention distribution in the Human Visual System. It is important to visual congition system; therefore, many computational visual saliency systems and applications have been built during the last three decades. This thesis analyses disadvantage of these models, and fouces on new saliency measure and description algorithms and applications. The main content and the original techniques are as followed:(1) Review the development of visual saliency, analyse of some classical model frameworks and the research results in recent years, and summarize disadvantage of the existing models, obtain the ideas of research on visual saliency.(2) Propose a salient edge extraction method based on spatial facilitation and surround inhibition. The method can effectively suppress the non-salient edges from texture and background. In this way coherent spatial configurations and region boundaries could stand out from their surround. Finally saliency of edges is measured by the Gabor energy of the edges by combining with edge attribute model. This method highlights the important edges in the imges (such as the contours, important texture edge, etc.), and weaken other edges, like the Human Visual System does. From this measure to salient edges, it is easy to obtain the boundaries of a image and the experiments demonstrate the method is as good as the learning models of boundaries extraction.(3) Propose a novel saliency description method based on object information. Different of traditional description method, this method describes saliency by combining the saliency map and salient object region information. We optimize a saliency map from a location medel by region information and Experiments demonstrate that the idea is right. Moreover, because of combining with salient object region information, the model can simulation the analysis process of human visual for a natural scene.(4) Propose a saliency model based on superpixels and use it in salient object segmentation. The saliency model combines the features extracted from superpixels by a global-local framework and avoids the inconsistent problem of saliency measure. Experiments show that this algorithm is better to measure the region saliency than others. In salient object segmentation experiment, the model outperformed others in salient object visual integrity.(5) Propose a fast algorithm to detect weak targets in remote sensing image based on visual saliency. The algorithm combines the spatial residual model and frequency-tuned model to obtain the ROI (ROI:regions of interest). Moreover, we suggest a method of dynamic background modeling to automatic segmentation of ROI, and remove the false regions using morphology filters. Experimental results show that the algorithm improved much in all respects, compared to classic visual significant model.