Hierarchy Semantic Management Of Image Based On Saliency Detection

Automatic image classification and management is one of the hottest demands as wellas one of the difficulties in intelligent processing in the digital world. Based on themechanism that the semantic information can be quickly extracted according to humanvisual system’s perception to the external environment, the research on hierarchy semanticimage management based on saliency detection is a meanningful attempt to improve digitalimage management and query. In this paper, making the static image as the research object,we study the method for clarity judging, saliency detection model and hierarchy semanticmanagement of image based on saliency detection. Our researches are as follows:1) According to the characteristic of human visual selective mechanism that humanalways focusing on the clear regions, we propose a decision rule to judge whether there isclarity difference between sub-regions, which is based on the dispersion of the highfrequency map(HFM). We construct HFM for each image because it can accurately reflectthe degree of high frequency component’s dispersion, and thus can quickly and accuratelyjudge whether there exists clarity difference in the image. At the same time, we present aneffective blur region inhibition algorithm for those images that exists clarity difference,which is based on the gradient method and wavelet transform. The process of judging andinhibiting can effectively simulate the mechanism of human’s selection for clarity, and thushas greate influence on saliency detection and image cognition.2) We develop a saliency detection model which can be useful to the problems ofcontent missing in image with large scale foreground and false detection in image withcomplex background when detecting salient regions with the existing models. The modelfirstly adopted the blur region inhibition to reduce false detection to some extent, and thencombined the low and high level saliency features. In low level features, it merged thesaliency information both in space and in frequency domain while in high level features, it took the face saliency into consideration. After some fusion and center-surroundcomputations, the final saliency map can be achieved. Experimental results show that theproposed model can better solve the present problems with considerable high efficiency.Salient regions exhibit human’s saliency cognition of the image and have the guiding effecton image cognition and management.3) We extended the applications of saliency detection model to the area of imageorganization and management. For the problem of time-consuming， labor-intensive andlow efficiency in image management, an automatic hierarchy classifying model for imageset was constructed. Based on the traditional annotation model and saliency detectionmodel, our model firstly established a hierarchy semantic label tree(HSLT) for imagehierarchy annotation, according to which each image can be put in to corresponding files inthe hierarchy file model. The model can vividly simulate the process of human thinking inimage management and can achieve better effect.