一种图像局部特征的语义提取方法

A Method for Extracting Image Semantics of Local Features

本文提出了一种基于期望最大化(EM)算法的局部图像特征的语义提取方法。首先提取图像的局部图像特征,统计特征在视觉词汇本中的出现频率,将图像表示成词袋模型;引入文本分析中的潜在语义分析技术建立从低层图像特征到高层图像语义之间的映射模型;然后利用EM算法拟合概率模型,得到图像局部特征的潜在语义概率分布;最后利用该模型提取出的图像在潜在语义上的分布来进行图像分析和理解。与其他基于语义的图像理解方法相比,本文方法不需要手工标注,以无监督的方式直接从图像低层特征中发掘图像的局部潜在语义,既求得了局部语义信息,又获得了局部语义的空间分布特性,因而能更好地对场景建模。为验证本文算法获取语义的有效性,在15类场景图像上进行了实验,实验结果表明,该方法取得了良好的分类准确率。

A method for extracting image semantics of local features is presented based on the Expectation Maximization algorithrn（EM）. The local image features are first extracted and the visual words in codebook are used to describe every feature, and then the semantic model mapping from low-level image features to high-level image semantics is achieved by using probabilistic latent semantic analysis. The latent semantic probability distribution is calculated for local features, and their spatial distribution in image is calculated using the Expectation-Maximization algorithm. Finally, this semantic probability distribution is used to image analysis and understanding. Compared to other semantic-based image understanding methods, the proposed method extract local latent semantics directly, which does not require manual annotation. It not only obtains the local semantic information, but also receives the distribution of semantic space. And thus it is better to model the scenes. The experimental results show that this method has satisfactory classification performances on a large set of 15category scenes.