一种基于Corr-LDA-ITD模型利用类标促进标注的方法研究

Research on the method of class label promoting the annotation based on Corr-LDA-ITD model

Corr-LDA-ITD模型较Corr-LDA模型具有更高的文本图像关联性,并且类标信息可以一定程度影响标注的准确性。基于此,文章首先利用Corr-LDA-ITD模型对不同类别的图像进行训练,然后用每类图像训练出的模型对测试图像进行测试,选出log似然值最大的标注模型对图像进行标注。在2个真实图像集(Labelme和Uiuc-sport)上的实验结果表明该方法的标注效果较好。

Corr-LDA-ITD model has higher text image relevance than Corr-LDA model,and the label information can affect the accuracy of the annotation certainly.Therefore,this paper uses the Corr-LDA-ITD model to be trained on different categories firstly,then uses the trained models to test the test images,selects the annotation model of the largest log likelihood to annotate the images.The experimental results on two datasets(Labelme and Uiuc-sport)show that the annotation performance of the paper is better than the comparison method.