一种基于概率产生式特征学习的图像检索方法

Image retrieval based on probabilistic generative feature learing

提出了一种基于概率产生式特征学习的图像检索方法,该方法采用高斯混合模型对数据分布建模,并基于贝叶斯推断的方式从模型中学习出表征图像的特征映射,该特征映射充分利用了数据分布信息。具体来说,它编码了概率产生式模型的模型参数、隐变量和观测样本等有益于图像检索的高层语义信息。最后,基于学习出的特征映射构建核函数衡量图像间的相似度,完成图像检索。在两个标准图像数据库的实验结果验证了本文方法的有效性。

We propose an image retrieval approach based on probabilistic generative feature learning, which uses Gaussian mixture models to model data distribution, and learns image features from the generative model based on Bayes inference. The learned fea- ture mappings fully consider data distribution information. Specifically they encode high-level semantic information for image re- trieval over model parameters, hidden variable and observed samples. At last, this paper constructs similarity function to measure similarities between images based on the learned feature mappings. The proposed approach is evaluated on two standard image datasets. The experimental results validate the effectiveness of the approach.