提出了一种自适应的Web图像语义自动标注方法:首先利用Web标签资源自动获取训练数据;然后通过带约束的分段惩罚加权回归模型将关联文本权重分布自适应学习和先验知识约束有机地结合在一起,实现Web图像语义的自动标注.在4 000幅从Web获得的图像数据集上的实验结果验证了该文自动获取训练集方法以及Web图像语义标注方法的有效性.

This paper proposes a novel adaptive model for Web image semantic automatic annotation. First, the model automatically collects training image data by exploring the associated textual data and the social tagging data of Web images, such as the Flickr’s Related Tags. Then, using a newly constrained piecewise penalty weighted regression to combine the adaptive estimation of the weight distribution of associated texts and the prior knowledge constrain together and implement the Web image semantic annotation. The proposed training data auto-generation methods and Web image annotation approaches are tested on a real-world Web image data set and promising results are achieved.

Supported by the National Natural Science Foundation of China under Grant Nos.60403018, 60773077, 90818023 (国家自然科学基金); the National Basic Research Program of China under Grant No.2005CB321905 (国家重点基础研究发展计划(973))