Image Annotation Based On Feature Reconstruction And Semantic Correlation Calibration

Image automatic annotation is an active research focus in the field of image processing. Improving the accuracy of image annotation is significant to improve the management effectiveness of the image data. The relative technical methods have applied into many other filed. Many image automatic annotation algorithms have been proposed in recent years, and most of existing algorithms contain three categories:probabilistic methods; topic methods and classification methods.However, as the fast growing of databases, more limitations have emerged urgently in this field. The first aspect is the semantic gap problem. The second is that the word cannot be connected to the target area directly. In order to solve these problems, many improved algorithms have been proposed. Up to now, SML is one of best methods which is based on the classification technology. Besides, many other popular methods have been raised, such as CMRM, CRM and MBRM, which are mainly based on probabilistic methods. However, the disadvantages of those algorithms is obvious.In order to deal with the semantic gap problem, we proposed a weighted feature reconstruction algorithm. The method can effectively enhance the visual content that contained in the feature vector, and apply the SVM to train the image data.Experiments drawn the conclusion that the proposed algorithms are significantly effective than some parallel methods.For further improvement of the performance, we proposed a semantic correlation calibration method to improve the quality of annotation results. Some redundant or wrong words have been eliminated through this method. The experimental results shown that, the calibration method based on semantic correlation obviously improved the image annotation performance.The above methods are tested on the Corel5 k, which showed that it outperformed some traditional algorithms.