Web-based Approach For Automatic Image Annotation

The content-based image retrieval leads to the "semantic gap" between imagelow-level features and the richness of human semantics. However, the effective way toreduce the "semantic gap" is still unavailable. With the development of the Web2.0,there are more and more users active on the Web. While uploading images, most ofusers usually provide some annotations or corresponding description text which is veryimportant for mining high-level semantic features. However, most of previousresearches mainly concern how to design good machine learning method to annotateimages. There is still no effective approach for making full use of the textualinformation.In order to avoid complex learning process and process the large number ofinternet images, this thesis designs the Web-based approach for automatic imageannotation. The main content of this thesis is as follows:Firstly, the image processing techniques and fast-and-accurate K-means are used toaccomplish the text representation of the image, which is the basis of the followingresearch work. And some corresponding text processing techniques are applied toprocess images.Secondly, in order to analysis the advantage and disadvantage of the traditionalmachine learning method, the state of art multi-label classification algorithm MLKNNand MFoM are implemented to annotate the image.Finally, the framework of the Web-based automatic image annotation approach anddetailed algorithm of the method are designed which combines the text representation ofimages, content-based image retrieval, natural language processing techniques.Experimental results show that Web-based approach for automatic imageannotation is much better than the machine-learning-based approach on both precisionand recall, when the scale of the image and label set is not very small. The Web-basedapproach not only avoids a large number of parameters optimization and complexlearning process but also the heavy manual work of annotating image training set. Italso can mine the relatively rich semantic labels from the image set whose label isrelatively sparse.