Image Annotation Research Based On Salient Region

Because of the rapid development of the Internet and the continuous increase in mobile terminal devices, the number of images rose sharply. Large image data will put forward higher demand for image processing technology inevitably. It has become an urgent and challenging task that how to find the target image quickly and efficiently from the huge image database. Image annotation technology is one of the key technology of digital image semantic text information, which is widely used in all aspects of digital image processing.In the field of image annotation, annotation performance maybe can get better results with the help of other related disciplines besides the existing algorithms. In biology, the human visual system can extract the semantic information of the image quickly and accurately. Therefore, it will be a beneficial attempt to apply the human visual attention to the research of image annotation.The main contents of this paper are as follows:(1) This paper analyzes and summaries the salient region detection methods have been proposed. Besides, it proposes a new method based on visual attention model according to the area of each region for the whole image and the variance of surrounding areas'brightness. The model is designed to simulate the way people get meaningful information from the image and set about the low-level feature. It belongs to the bottom-up detection method;(2) The key technologies of image annotation in recent years are analyzed and studied.The visual attention model is applied to image annotation, and a method based on visual attention model and k-NN clustering algorithm is proposed. Firstly, salient region is obtained by visual attention model,then the features of salient region and non salient region are extracted and fused, finally k-NN clustering algorithm are combined to image annotation. The results show that there is a great relationship between the results and the proportion of significant and non significant areas;(3) The salient regions of each image are separated based on the visual attention model proposed in this paper. Then the salient regions and non salient regions are marked,and the relationship between salient regions and non salient regions is analyzed. So a method based on support vector machine and word correlation is proposed. This method can eliminate the influence of non salient regions for the salient regions. The experimental results show that the method is feasible and effective.