Automatic Image Annotation Research Based On Extreme Learning Machine

Today is the era of explosive growth in the image.It is urgent to advance the research of images so that the capability of images management and understanding could keep consistent with the increasement of images. Image annotation can not only provide a more comfortable service represented by web image search for common users, but also will bring revolutionary changes to the organization,indexing,management of the image data, even promoting the image understanding and the entire field of computer vision progress. In addition, image annotation is also important in commercial applications, digital libraries, military, biomedical, network monitoring, and other aspects of national security.In this paper,I first introduced some basic knowledge of image annotation,including image segmentation, feature extraction, annotation model and evaluation standards.There are two parts in image segmentation, including fixed partition and N-Cut algorithm. About the feature extraction, I introduced color, texture, shape and other characteristics. The main evaluation standard are recall and precision.This paper focuses on the image annotation models,image annotation based on classification,image annotation model based on probability, and the image annotation model based on the theme. The first two models attempt to find a direct relationship between the image(or image region) and the annotation words,but the image annotation model based on the theme by introducing the concept of the potential topics to establish contact with the high-level semantic words between low-level visual features enable automatic image annotation.Image annotation method based on machine learning has been widely used and developed currently. Therefore the selecting of the learner model is very important. This article advanced a research focused on how to improve the eff ectiveness and efficiency of image annotation, and proposed an algorithm on image annotation based on elm.The feature vector of this algorithm is made up of color feature, texture feature and SIFT feature, so it is a compound characteristic vector, can reflect the characteristics of image fully, and improve the efficiency of retrieval effectively. In the model training period, we use elm algorithm because of its simple structure, fast learning and good generalization performance. With the development of ELM, theoretical research and practical have shown that ELM not only can apply to neural network,but also have agood performance in the application of multi-classification. The results of this experiment show that the algorithm not only has a huge cost advantage in terms of time, also its good generalization performance improved the accuracy of the algorithm, the combined effect of the two aspect make the image annotation performance has been effectively improved.Finally, design and implement a image retrieval demonstration system.