The Research Of Automatic Annotation Method For Natural Scene Image

With the of the advancement of technology, the continuous development Internet and the popularity of the digital products, more and more non-textual information, such as image appears in people's life. An efficient image annotation and retrieve method is highest needed, for a large number of image information management. Early image annotation often manually put up some keyword for an image to describe the content of the image. And then make index on these words in order to convenient for the retrieving of the image. But the quantities of image increased very fast, almost exponential, manual image annotation became expensive,labor intensive and easy bring errors. So people have been great desired in labeling image in an automatic ways. Besides automatic image annotation is significance for image understanding and retrieve of web image, so it becomes the new hot research topic in recent years.On the basis of serious summarizing some image annotation methods, we propose an efficient method for image annotation, which is the method of CMRM image annotation based on inter-word correlation. In this thesis, image annotation is divided into two parts. First use the existed CMRM to make basis annotation. Then the semantic relationship between annotations is analyzed. The correlations between keyword is picked up, and stored in a matrix which named inter-word correlation matrix. This matrix describes the semantic correlations. Finally, the inter-word correlation matrix is added to the initial labeling matrix by graph learning algorithm, making the correlation between words to be spread between the various keywords and then improve the results of the annotation.In the end of the thesis, in order to testing the method proposed in this thesis, experiment has been done on some natural scene image which comes out of the Corel image database. The experiment result indicate that the method obtain good performance on automatic testing image annotation. Compared with the CMRM, our method gets higher recall and precision. Compared with the Co-occurrence Model and Machine Translation Model, the recall improves 3 times and 1.5 times and the precision improves 5.5 times and 2.7 times. Besides of these, the number of keyword which can be correct retrieved also improves 3.5 times and 1.3 times. All of these could testify the efficient of our method in automatic image annotation.