【作者】 陈恋； 周向东； 叶剑烨； 谢祖铭； 张亮； 施伯乐；

【机构】 复旦大学计算机与信息技术系；

【摘要】 <正>1引言早期的图像数据库系统通过对图像进行人工标注实现图像检索功能。然而,传统人工标注存在着效率较低,语义标注的一致性难以保证等问题,图像语义的自动标注正成为新的研究热点。对图像进行离散化处理,然后应用文本信息处更多还原

【Abstract】 We propose a novel stratification-based automatic image annotation approach which can be used in multimedia database.In our approach,we first segment images into regions.Each region is represented by a feature vector. Then we employ a minimal-entropy based feature discretization method to get discretized visual features.We propose a concept of image document model,looking at the structure of images from a specific view and providing a more reasonable expression of images.Finally,a method based on classic vector model is used to enable automatic image annotation according to the similarity of image documents.Our experiment using real datasets shows that our model outperforms the Co-occurrence Model,the Translation Model and Cross-Media Relevance Model significantly. Our model performs better than the state-of-the-art Continuous-space Relevance Model(CRM) in recall and it is nearly 400 times faster.更多还原