Combining Visual And Textual Information For Automatic Image Annotation

Automatic image annotation is extensively concerned in fields of image retrieval. Itsapplication has ranged over many areas, including the establishment and management ofdigital libraries, automatic medical image annotation, retrieval and management of digitalphotos, and so on. Recently, combining visual and textual information for automatic imageannotation is the development direction of of automatic image annotation. Based on theidea of machine learning approach, we research number of key issues about combiningvision and text information for automatic image annotation, including the combiningsemantic relevance measure, and the model of combining vision and textual information.An approach to combining semantic relevance measure is proposed. It is combiningsemantic relevance measure based on neural network. The approach describes therelationship among the semantic relevance measures, by making use of the characteristicthat 3 Layer neural network could express every complex nonlinear relationship. It uses theglobal optimization feature of particle swarm optimization algorithm to adjust the weightsof neural network. During neural network learning, the process of automatic imageannotation is seen as a whole. The way of learning is cross-learning mode, that is, particleswarm optimization of neural networks and the model of combining vision and textualinformation are crossing each other to learn, until the the structure of automatic imageannotation reaches a steady state. In the Corel5K image library for the experimentalverification, the result of the 10 candidate annotation words shows that combining visionand text information for automatic image annotation increases by about 23.2% thanautomatic image annotation based on visual information.An improving approach to the model of combining vision and textual information isproposed. The model adopts particle swarm optimization algorithm to optimize the processof image vision and textual information. It makes number of parameters in the model tolearn automatically. This way of automatic learning mode can avoid the uncertainty bysetting parameters artificially and reduce the tedious nature of manual operation. Byreducing the interference of human factors, the reliability of algorithm is improved. In theCorel5K image library for the experimental verification, the result of the 10 candidate annotation words shows that combining vision and text information for automatic imageannotation increases by about 23.6% than automatic image annotation based on visualinformation.