Saliency Detection Method Research Integrating Global And Local Information

For more than a decade,the endless stream of information technology in computer vision is the key to solving image data redundancy.Saliency detection is very important research direction in computer vision,and application is very extensive.It originated from the human visual attention mechanism,and aims to find out the most interesting area in human eyes.So that this can reduces the redundancy in images and video,improve the efficiency of the subsequent image processing.As a result,saliency detection research has far-reaching effects on many tasks such as image recognition,image segmentation,video behavior recognition.In the saliency detection research,the integration of global information and local information has achieved the good experimental results.However,these methods still have limitations.Aiming at its disadvantages of existing work,this paper presents a new saliency detection method.(1)A new similarity metric method is proposed to calculate the similarity between different regions of the image.In this method,feature weight is varied with image content.The similarity metric method is applied to the background-based map;it's easy to distinguish the foreground and the background.Then,we integrate three prior acknowledges to construct global saliency map which contains background-based map,center prior map and objectness map.(2)During the processing of obtaining foreground dictionary and background dictionary based on global saliency map,we develop a novel locality coding-based method to obtain more accurate saliency estimates.On the one hand,we propose a judge mechanism to gain more reliable dictionary.On the other hand,we creatively introduce codebook elements' reliability into calculate reconstruction errors.The saliency value of each super-pixel has obtained by reconstruction error with its local dictionary elements.Next,we get local saliency map.(3)Global saliency map can better highlight salient object and local saliency map can effectively suppress background noises.Therefore,we develop a novel integration mechanism to take both advantages of them.Different from the previous work integrating global and local information directly,proposed integration method endow them with different weights according to the different performance of global and local saliency map.Finally,an integration mechanism is proposed to further improve performance.Proposed methods are compared with other 13 saliency detection methods in four datasets,experimental results demonstrate the superiority of proposed algorithms.Meanwhile,parameters and each component about proposed algorithms are also deeply analyzed.