一种基于空间金字塔模型的兴趣点对自相关的图像特征

摘要

This paper presents a new approach to improve the traditional bag-of-visual-word model for scene categorization. Tradi-tional model considers images as a histogram of the occurrence rate of interest regions. In this approach, the spatial distribution of code words is incorporated to approximate the image geometric information. This works by improving the traditional codeword his-togram with calculating the spatial distance between pair wise interest regions. It combines the approach with spatial pyramid matching algorithm to consider global geometric information and strengthen its ability to represent the image content. Experiment results on a public dataset show that the combination with spatial pyramid matching increases the accuracy and improves effective-ness for categorization.%针对传统视觉词袋模型只考虑兴趣点出现的频率而忽略了局部特征空间信息的问题，提出了一种基于空间金字塔模型的新的图像特征。该特征在标准视觉词袋模型基础上，通过计算属于同一码字的兴趣点对之间的距离，加入了不同码字包含的兴趣点在图像上的空间分布。更结合空间金字塔模型，聚合不同分层过程中提取的特征，更大程度上考虑了空间信息，从而加强了特征对图像内容信息的表示能力。实验结果表明，与传统的词袋模型和金字塔模型相比，具有更高的精准度和分类性能。