Research On Image Automatic Marking And Retrieval Technology

Network technology as a new rapid development production mode not only meets people’svarious production needs, but it also provides user lots of image resources. In order to facilitateuser to achieve massive images according to their needs, the effective data processing ofappearing online images becomes an urgent problem. But there exists the understanding errorbetween the underlying visual features and semantic concepts generally. Artificial imageannotation has been difficult to meet users’ needs. In order to improve the image retrievaltechnology, automatic image annotation technology increasingly occupies a pivotal position.The automatic image annotation technology uses the correlation modeling technique withthe training set to establish the correlation from the underlying visual features to the high-levelsemantic information. Then, the unknown images can be automatically labeled according to theestablished correlation model. The correlation model method provides a fast extraction methodfor semantic concepts, which transforms the content based image retrieval into a simple textbased retrieval. This paper revolving around the correlation modeling technologies achieve theresearch about problems of the automatic image annotation modeling method and the imageretrieval technology fusing with semantic information. The specific researches are as follows:(1) The traditional FCM clustering algorithm has the problems that it is more dependent onthe clustering center and the clustering algorithm easily lead to the optimal solution which maybe of local problems. To solve these problems and improve the automatic annotationperformance, we use the gray histogram method and the innovation which increases theweighted values of the underlying image region feature improving the FCM algorithm to achievebetter clustering results. Then, using the Bayesian classifier model creates the correlation modelbetween the underlying visual features and the high-level semantic concepts for the markedtraining set. Lastly, we can achieve the image annotation for the unknown images according tothe maximum similarity between the testing image area and the training image area.Experimental results based on this technology illustrate that the improved annotation method issuperior to the traditional machine translation model, the CMRM model and the WFC model.(2) The definition method of correlation model is essential in the automatic imageannotation technology. A good correlation model enables better mapping from the low-levelimage features to the high-level semantic concept and further optimize the "semantic gap"problem. We can use the improved CMRM model to make the image annotation results better.When the user inputs an image to achieve the content-based image retrieval, we can use theimproved modeling method to achieve the semantic concept mapping from the input image. Experiments show that the retrieval performance of this method is superior to image retrievalmethods based on the TM model and the CMRM model.(3) Among the traditional automatic image annotation techniques, although the correlationmodel achieved the mapping from the low-level visual features of images to the high-levelsemantic information, they did not consider the impact of the semantic information for themapping. In other words, they ignored the impact of the correlation between semanticinformation. In this paper, we combine the semantic concepts and the weighted visual features toachieve the automatic image annotation. After that, we can use the symbiotic relationship amongconcepts to improve the annotation model and define the weighted visual feature for the imageretrieval sequence to achieve the sorting retrieval. Experimental results show our retrievalperformances are better than the CMRM model and the CRM model based retrieval method.