本文在扩展LDA（latent dirichlet allocation）的基础上提出了一种新的生成模型——基于类主题空间的潜在狄里克雷分布（CTS-LDA)用来实现自然图像场景分类。该方法不同于以往方法，它在训练时通过将图像场景类别信息引入模型推导过程中，产生各场景类的独立语义主题空间，使得每个场景类都有各自不同的主题空间，图像的最终语义表示采用与其类别相关的类主题集，是一种符合人类认知习惯的方法。以前所用的场景分类方法通常在得到图像主题表示后还需要依赖于其他分类器来完成场景分类，而CTS-LDA模型可以在分别计算图像在各类模型中的主题分布时，用最大似然法得出图像的类别信息。此外本文通过分析不同主题数对本模型性能的影响，得出了适用于本模型的最佳主题数。本文分别通过13，15等多类场景任务来检验模型的性能，实验证明该模型能够在不需要太多训练的情况下取得较好的性能。

The paper presents a novel model named Category Topic Simplex-Latent Dirichlet Allocation(CTS-LDA) based on extending LDA(Latent Dirichlet Allocation), which is used to learn and recognize natural scene category. Unlike previous work, our model can absorb category information in the model inference under learning process to produce its semantic simplex for each category scene. As a result, each category has its own semantic topic simplex, and each image can choose its simplex to denote, which is consistent with people’s cognitive pattern. In previous work, recognizing scene category task need to resort to additional classifier after getting images topic representation. Our method can use ML method to recognize image category during the same time of getting topic representation. Furthermore, we also analyze the influence of the topic size in our model, and infer the fittest result to produce the best performance. We investigate the classification performance under changed scene category tasks. The experiments have demonstrated that our model can perform better with less training data than other methods.