The Research Of Automatic Image Annotation

In recent years, with the rapid development of multimedia technology, the use ofdigital image is becoming more and more widespread in many areas, then leading to therapid number of imges rise. How to effectively manage these images and get theinteresting information for people has become one of the hottest research points.The automatic image sematic annotation learns the relationship of the annotationwords and images according to the training image set to get some rules, then it can getsome most suitable annotation words for an image, these words are the description of thehigh sematic of the image. Image annotation provides a method to pick up the sematic ofimages, then the sematic retrieval of images can be solved by the developed technologyof retrieval text.The joint probability between the the image and the positive bags of a word hasshown the importance of a word in an image, then it can be regarded as the weight ofword in this image. The weighted vocabulary relevance method(WVRM) analyzes theweight of an word in testing image, assumes the relation between words by multiplebonulay distribution, computs the joint probability between word and testing image,finally gets the probability of a word in the testing image. Experimemts have shown thatWVRM gets an obvious progress than previous models, the speed of anntating an imagehas been cut down greatly and the model covers a great number of annotation words.Image annotation is a kind of superivsed machine learnning approaches, and itdepends on the training image set. It may get a better effectiveness by segmenting theimage set and choicing proper images as training images. A hierarchically automaticimage annotation method (HAIAM) partitions the words into some sematic clusters byanalyzing the similarity of each word. The whole images are divided into some sets interms of the corresponding relationship between the image and annotation words, after dividing each set has a common semtic topic. When a testing image requires to annotate,the method judges the topic of the testing image according to classifiers, and selects theimages of this topic as the training image set. After being pretreated, the image isannotated by some algorithms in detail. Experiments have shown that the HAIAM canget better results than previous models, especially in the precision of the annotationwords.