基于核密度估计的图像自动标注方法

Image Automatic Annotation Method Based on Kernel Density Estimation

提出一种基于概率模型的图像自动语义标注方法,将图片自动标注看作一个多类分类问题,通过无参数的核密度估计,实现用含有共同标注词的图片组估计视觉特征和相应标注词之间关系的机制。选取表达能力较好的基于CPAM的视觉特征,无须对图像进行语义分割处理,有效提高核密度估计的效率。在基准数据集上进行实验,结果表明,该模型能够获得比当前其他相关方法更好的标注性能。

A novel method for automatic image annotation is presented. The new model is based on a probabilistic formulation which poses annotation as a multi-class classification problem. It tries to estimate the conrrelation between visual features and semantic labels by using the groups of images that share the same associated labels through kernel density estimation. In addition, CPAM-based visual features are introduced to improve the efficiency of kernel density estimation without requiring prior image semantic segmentation. Experiments on the benchmark data set show this model achieves higher accuracy than the previously published results.