Automatic Image Annotation

With the development of Internet, the quantities of various images increasing rapidly, how to index and manage the images efficiently from the huge image database has become a very urgent issue to be solved. Now, a lot of Internet companies have their own search engines, but most of them use the images retrieval technology base on text, and this technology need retrieval by human, that’s waste of time. Although it’s a long time for the research on images retrieval based on the color, texture, shape in the image low-level feature for similarity matching,restricting by the semantic gap, there are still some difficulties on the retrieval accuracy.This paper presents a semantic automatic annotation method of multi-instances image based on joint feature selection. it can mark the test images’ labels automatically by training the semantic of images. Base on this research, this paper extend the research on image annotation, and put forward images’ labels estimation of distribution algorithms based on maximum entropy model. So we can get both the labels of test images and the distribution of labels. The main research of the paper includes:1) As the inaccuracy of semantic annotation, this paper has researched the application of multi-instances study on image. The first is the instances of image to be divided and the underlying features to be extracted. Then features and label of these instances are input to the learning of algorithm model. The multi-instance study can solve the automatic annotation issues both single-label and multi-label images.2) As the finiteness of image labels’ quantities, this paper has researched the application of semi-supervised learning on image annotation. Selecting both the labeled samples and no-labeled samples, then learning them as the training samples. By using the graph Laplacian to solve the semi-supervised learning issues.The first is building the weighted graph. Then we will get the Laplacian operator.Finally we put them into the Manifold Regularization to optimize the objective function.3) This paper designed the multi-instance multi-label image automatic annotation arithmetic based on feature selection. On the basis of multi-sample study,this paper has a semi-supervised learning on the database, and use the manifold learning method reduced the dimensions on the high feature dimension data.Meanwhile, this paper built the linear regression model combining the sparse representation, optimized and got the multi-instance multi-label automatic annotation arithmetic based on feature selection. After getting the predicted value by training,and getting the label by determining the predicted value using adaptive threshold method, and then got the accuracy of image annotation.4) This paper has extended the image annotation, and put forward images’ labels estimation of distribution algorithms based on maximum entropy model. It can get both the images’ labels and the distribution of each image, and get the proportion of each label in each image. Users can select the image that is the most proportion in the searching labels, it can improve the accuracy and rationality of image index. Inspired by the labels distribution learning, this paper use the multi-sample multi-label image automatic annotation arithmetic in labels distribution learning study, and get the good result.