基于有监督Topic Model的图像分类

Picture Classification Based on Supervised Topic Model

近年来,以LDA为代表的话题模型在图像和文本处理中均得到了广泛的应用。与传统的机器学习方法相比,LDA模型具有参数少,表达能力强等优点,同时作为一种生成模型,它可以有效模拟人类学习的方式,便利地加入先验知识。有监督的LDA模型则将生成模型与判别模型结合在一起,是一种通用的分类方法。Dense-SIFT特征被作为底层特征,在词袋模型的框架下,以k-means算法构建词典,用有监督的LDA模型训练,并在通用的图像数据集上进行评测,根据评测结果证明其在图像分类任务中具有很好的性能。

In recent years, Topic models, which are represented by LDA, have been widely used in both image processing and text processing tasks. Compared with traditional machine learning methods, LDA model has less parameters, and a stronger ability to capture deep structure of data. Also, as a kind of generative model, LDA model can simulate the learning process of human, and is able to integrate priori knowledge easily. Supervised LDA （sLDA）, which combines generative process and discriminative process, is a common model for classiifcation. Dense-SIFT is used as low-level features and k-means algorithm is applied to construct dictionary, then a classiifcation model is trained using sLDA, we evaluate this method on a populardataset, under the framework of bag-of-words, which proves that it works well on Image classiifcation tasks.