基于LDA主题模型的图像场景识别方法

摘要

针对传统潜在狄利克雷分布(latent Dirichlet allocation,LDA)主题模型在进行图像场景识别时存在聚类方法效率低以及不能有效利用图像主要特征的问题,提出改进图像场景识别模型的方法.采用K-Means++聚类算法生成视觉单词,使用加权统计直方图完成图像表示,通过引入特征函数加强重要特征在分类识别中的作用,提出有特征函数的潜在狄利克雷分布(featured latent Dirichlet allocation,FLDA)主题模型.实验结果表明,对比于改进前的模型,该模型可缩短执行时间并提高识别准确率.%To solve the problem that clustering methods are inefficient and the main features can not be effectively used in the tra-ditional latent Dirichlet distribution (latent Dirichlet allocation,LDA)topic model for image scene recognition,the traditional model for scene classification and identification was improved.To generate visual words,K-Means++ clustering algorithm was used,and to complete image representation,weighted histogram was used.A feature function was added to strengthen the im-portance of the important feature in the classification and recognition.Latent Dirichlet distribution (featured latent Dirichlet allo-cation,FLDA)model with feature fuction was proposed.The experimental results show that compared to the previous model, the proposed model can improve the accuracy while shortening the execution time.