目前主题回归多模式潜在狄利克雷分配（tr-mmLDA），一个新颖的统计主题模型的图像和视频注释的任务。 在我们的新注释模型的核心是一种新颖的潜变量回归方法来捕获图像或视频特征和注释文本之间的相关性。 我们的方法不是在两个数据模态之间共享一组潜在主题，如在对应关系LDA的公式中，我们的方法引入了回归模块来关联两组主题，其捕获更一般的关联形式，并允许主题的数量 2个数据模态不同。 我们证明tr-mmLDA对2个标准注释数据集的功率：一个5000图像子集的COREL和一个2687图像的LabelMe数据集。 所提出的关联模型示出了与标题茫然性所测量的对应LDA相比改进的性能。-Currently topic regression multi-mode potential Dirichlet Allocation (tr-mmLDA), image and video annotation theme of a novel statistical model task. At the heart of our new annotation model is a novel method to capture latent variable regression correlation of images or video features and comment text. Our approach is not in the data between the two modes share a set of potential topics, such as correspondence between the LDA formula, our method introduces two themes related to the return module, which captures more general form of association, and allow the data relating to the number of two different modes. We prove tr-mmLDA of two standard annotation datasets power: a 5,000-image data sets COREL sub LabelMe and a 2687 image set. Associated with the proposed model shows a title corresponding to the measured loss of improved performance compared to LDA.