中文摘要：

提出一种基于类别约束的主题模型用于实现场景分类.不同于现有方法,本文将图像场景类别信息引入模型参数推导过程中,采用与其类别相关的类主题集描述图像的语义内容.针对各场景类图像中潜在主题数量变化,提出了一种ATS-LDA（自适应主题数的潜在狄里克雷分布）模型实现中层语义的建模算法.该模型依据各场景类训练样本关于视觉词语表示的变化估计所需主题数,体现了各类场景中间语义的繁简变化.根据各类模型下的图像概率分布,采用最大似然估计实现测试样本的场景语义分类.改变了现有主题模型需要依赖于其它分类器完成场景分类的现状.通过多个图像数据集分类任务证明该模型能够在不需要太多训练的情况下取得较好地性能.

英文摘要：

A novel topic model has been proposed to classify image scene,which is trained in the limitation of category constraint.Different from previous methods,the image category label has been used to infer model parameters,and different topic set related to category has been built to describe the image semantic information.According to the variation of topic among all images,ATS-LDA（Adaptive Topic Size-Latent Dirichlet Allocation） has been proposed to learn image scene,in which the size of topic set for each category scene is adjusted adaptively based on its semantic content.Our model makes use of ML（Maximum Likelihood） to compute the likelihood in all category models,and recognize the category label that the image belongs to.Some experiments have been done to demonstrate good performance of our model.