Research On Image Semantic Automatic Annotation Based Scene Classification And Retrieval

Image semantic automatic annotation is an effective way to improve the performance of image retrieval. If using traditional text-based modes to find the desired image,we need detailed manual annotate each image in first. But the method of manual annotate is very time-consuming and labor-intensive,at the same time,it is error-prone. For a variety of problems to the image semantic automatic annotation methods, this paper propose a method to improve the annotation performance,focusing on the study of the scene semantic classification and annotation algorithm.Using semantic automatic annotation can learn and obtain the relational model automatically,which is between bottom features and top semantic of the image. Using this model to annotated the uncommented images can compensate the "semantic gap" effectively. But the current annotation method is applicable to the specific field of image retrieval generally; The accuracy of the training set is annotated affect the performance of machine learning seriously; And It' s difficult to identify the scene semantics, especially high-level semantics.For the problem of accuracy in annotation, the paper use a image automatic annotation method of support vector data description. The method establishes similar-sphere for each type class separately, which can reduce the scale of solving problem and the complexity of the algorithm.Against to situation of the scene annotated ineffective, this paper presents a method of image semantic annotation, which combines PLSA scene classification with support vector data description. It uses the method of PLSA to find the potential semantic theme of image, and the method of support vector data description to construct classifier. Then use the classifier to annotated the unknown image. Compare to the support vector data description method, experiments show this method not only improves annotated effect, but also improve the retrieval performance. From the harmonic-mean value(T) of the comparative experiments, the results obtained by this method are higher than 60% basically, while the results of SVM and K-nearest method are lower than 50%, It explains the method has better label performance; For recognition of the scene semantic, the value of the T that obtained by this method is up to 74%,which is much higher than the other two methods. And it also proves thatthis method can improve the capabilities of scene semantic annotation effectively.