Research On Automatic Image Annotation Based On Visual Attention Mechanism And Support Vector Machine

In recent decades, with the rapidly development of digital technology and Internet technology, a large amount of digital images appear on the Internet every day, as a consequence, people realize that it is becoming increasingly difficult to find the pictures they need quickly and accurately on the Internet, so image retrieval technology is appeared to solve the problems of people searching pictures. But, the professional image search tools is burdensome for most users, it usually need users to provide a similar image, which increase the users’work. Automatic image annotation technology turns this process to text retrieval, users only give a key word, and the system will return the images related to the word. This is consistent with the majority of the user’s habit, what’s more, text search technology is more mature than image search technology. So automatic image annotation has become an important research topic in the field of image search,, as long as the image is accurate labeled, it can provide the users of satisfy results.The main factors affecting the annotation results contain the following aspects:1. Accurate image segmentation, which means identify the images’meaningful areas and regions of interest.2. The method of extracting and representing the image features (color, texture, shape, spatial etc.).3. The model of training images and words (mainly contains classification based model and probabilistic model).4. Refine the results when the annotation words are given.Currently, the image annotation algorithms often extract the image features as a whole, or seldom considering of the different importance of each region in the image. To solve these problems, we propose an image annotation algorithm based on the vision attention mechanism and support vector machine, First, in the pre-process, it separates the image into two parts:the salient region and the rests, by using HC algorithm, and compute their feature vectors respectively. In the process of annotation, first we obtain the overall annotated words of an image by using its weighted feature vector, annotated model and words probabilities model, then extract the salient words from the overall annotated words, and feedback the annotation result to refresh the words probabilities model. Consequently enhance the precision of image annotation. We do experiment on the Corel5k image sets, compare the annotation results with original SVM-based annotation results, and the result is satisfied.