Automatic Image Annotation Of Semi-Supervised Learning Based On Voronoi Graph

With the development of image recognition technology and content-based image retrieval technology, The automatic image annotation technology has received unprecedented attention.So it gradually becomes an important research topic in the field of image processing and pattern recognition.Automatic image annotation method is proposed to solve the problem of the traditional text and content-based image retrieval systems, such as the huge workload and the "semantic gap".The method firstly construct a relational model between image visual feature space and image semantic concept space based on labeled and unlabeled image set.Then it let the computer to study the model according to the existing information automatically. Finally, the task of labeling the unlabeled images is realized. However,how to effectively extract and used the low-level visual features of describing accurately image semantics, and how to effectively construction a fast and excellent information channel between low-level image features and high-level image semantic, and how to optimize the result of tagging?These questions will be the key to realize the automatic image annotation and are also very challenging task for many researchers. In this paper, To solve these problems, the automatic image annotation method based on graph is an important foothold. The theoretical basis and implementation process of the method are further studied, and the improved scheme is given.The first and most important task of automatic image annotation is to extract the low-level visual features of the image. This paper focuses on two kinds of color and texture feature description method based on color histogram and Gabor wavelet transform. In view of the problem that the traditional feature fusion method of fixed weight does not take into account the internal relations of the feature vectors and the need for a lot of experiments to determine the weights, a feature fusion method based on the weight matrix is proposed. And experiments show that the method can be used to better image retrieval performance.In semi-supervised learning method based on traditional model of graph, In view of the problem that the similarity graph with images as nodes did not fully take into account the data from a subset of data structure distribution, this paper puts forward a semi-supervised learning method voronoi-based k order adjacency.Based on the principle and mechanism of semi supervised learning method voronoi-based k order adjacency,a semi-supervised learning automatic image annotation method based on Voronoi graph is proposed in this study(VGSSL). This method fully takes into account the natural superiority of the Voronoi map in the expression of the influence region of the space target, and goodly combine the distribution information of the image data points in the feature space. Finally, this method achieved the transfering of annotation words of marked images from marked images to unmarked images.so the task of labeling the unlabeled images is completed. Experimental results show that VGSSL can achieve higher performance compared to other methods, and the result is significantly improved.