Research On Detection Model Of Visual Saliency And Its Application On Augmented Reality

Human eyes can extract effective information quickly from complex scenes,thanks to a selective visual attention mechanism in human visual system. Manyresearchers have introduced visual attention mechanism into image processing, whichplays an important role in image understanding and analysis. In recent years, saliencyanalysis has been successfully applied in object detection, recognition, imagecompression and other areas. In this paper, we propose a new approach for saliencydetection, and discuss the application of saliency detection in Augmented Reality (AR)system.Augmented Reality is a developing technology, which integrates graphictechniques, visualization techniques and sensor techniques. AR realization can mixvirtual objects with real backgrounds, and vise verse. It enhances the visualexperience of real scene. Augmented Reality is one of the most exciting technologiesand will provide powerful means for human intelligence extension. It will impact onhuman’s lifestyles profoundly. By combining AR and saliency detection technologies,context-aware services become possible, which can provide people-orientedintelligent services.The paper firstly analyzes several classical saliency detection models. Then wepropose a saliency detection model, using the feature of intensity and color. In thismodel, we calculate the contrast between image patches globally to imitate thereceptive field mechanism of visual nerve. It is named GPD model. With the model, agray image will be generated, reflecting the saliency of each pixel.In our saliency detection model, we also introduce the feature of multi-scalecontrast, which contains information about contour information of salient region. Wecreate Gaussian multi-resolution pyramid model to simulate the awareness feature ofhuman visual system. Based on the pyramid model, we can easily extract edge featureof salient region. Then, with the combination of GPD map and multi-scale contrastfeature, a saliency map is generated. The saliency map has clear outline of the salientregion and the interior region of salient region has high uniformity of luminancepattern. The paper also does research on how to extract salient region from saliency maps.We use two methods to segment the salient region. One method makes use ofthreshold analysis to binarize saliency map. The other is based on ConditionalRandom Fields. Large number of experiments verifies the effectiveness of ourmethods.Finally, we combine the saliency detection model and Augmented Reality, andcreate a simple program for primary fusion system and also design the basicframework of the advanced system.