中文摘要：

随着互联网的发展,网络图像指数般增长,图像理解技术变得日益重要.其中图像标注技术作为其关键技术得到广泛关注和研究.现有的图像标注技术大多是在图像层次上训练标签模型,忽略了图像区域之间的关系及其标签之间的关系.为了解决这个问题,文中提出了一种新的算法,结合区域之间的位置关系及其标签之间的共生关系辅助标注图像.具体而言,算法首先使用支持向量机对部分可确定区域赋予语义标签,然后利用区域位置关系帮助聚类标注未知区域.得到一幅图所有的区域标签后,我们提出两种模型对标签共生关系建模辅助修正标签集,一个是随机游走模型,另一个是条件随机场模型.最终算法输出每幅图像的文本标签集.在对图像集NUS WIDE的标注实验中显示,上述方法和单纯考虑区域关系的方法相比,标注效果和性能有了较好的改善,证实该方法是一种稳定、有效的标注算法.

英文摘要：

Nowadays,the amount of online images has grown explosively.This triggers the development of effective image understanding techniques.As a key technique,image annotation attracts broad attention.Most existing work learns models on the whole image for annotation task.However,these methods ignore the relationship between regions inside an image and the tag co occurrence relationships,which essentially limits the performance of image annotation.To tackle this issue,we propose a novel scheme that aims to exploit the region and tag co occurrence context for image annotation.First,we use the support vector machines （SVMs） trained on object categories to identify known and unknown regions.Second,spatial region context descriptor based clustering is used to annotate the unknown regions.Finally,two instantiation models,random walk and conditional random field （CRF）,are explored to refine the aggregated region tags for image annotation by utilizing tag co occurrence relationship.We conduct experiments on a subset of NUS WIDE.The results have demonstrated the effectiveness of our image annotation method.