基于区域空间与词汇加权的图像自动标注

Image Automatic Annotation Based on Weighted District Space and Vocabulary

图像自动标注是图像检索与图像理解中重要而又极具挑战性的问题.针对现有模型忽略了图像不同区域对图像整体贡献程度的差异性,提出了基于区域空间加权的标注方法,改善了图像的区域特征生成概率估计.此外,针对现有模型未考虑词汇本身重要性以及词汇分布对标注性能的影响,提出了基于词汇固定权值的标注方法、基于平滑词汇频率的标注方法以及基于词汇TF-IDF加权的标注方法,对词汇的生成概率估计部分进行了改进.综合以上区域空间改进与词汇改进,提出了WDVRM图像标注模型.通过在Corel数据库进行的实验,验证了WDVRM模型的有效性.

Image automatic annotation is a significant and challenging problem in image retrieval and image under-standing.Existing models ignored that different regions of images had different contributions to the overall images.So an annotation method based on weighted district space to improve the generation probability estimation of regional features of the images was proposed.On the other hand,existing model did not take into account the importance of vocabulary as well as vocabulary distribution which impacted the annotation performance.Three methods to overcome the above problems were proposed,including： fixed vocabulary weight method,smooth vocabulary frequency method and weighted vocabulary’s TF-IDF method.These methods can improve the generation probability estimation of vo-cabulary.By integrating all above improved methods of weighted district space and weighted vocabulary,WDVRM image annotation model were proposed.Experiments conducted on Corel datasets have verified that the WDVRM model is quite effective.