摘要： 针对传统视觉词袋模型对图像尺度变化较为敏感的缺点，提出一种基于改进视觉词袋模型的图像标注方法。该方法引入图像的多尺度空间信息，对图像进行多尺度变换并构建多尺度视觉词汇表，将图像表示为不同尺度特征，结合多核学习的方法优化各尺度特征的相应权重，获取特征表示。实验结果验证了该方法的有效性，其标注准确率比传统BoVW模型提高17.8%~25.7%。

关键词: 图像标注, 视觉词袋模型, 多尺度空间, 多尺度视觉词, 多核学习, 权重优化

Abstract: Aiming at overcoming the traditional Bag of Visual Word(BoVW) model’s sensitivity to image scale’s variation, this paper proposes an image annotation method based on improved BoVW model. It incorporates with multiple spaces information and transfers original images into multiple scale spaces and constructs multiple scale vocabularies. Images are represented as a family of feature histograms with different scale. Multiple kernel learning is introduced to optimize the histograms weights of different scale in order to acquire discriminative classifying power. Experimental results prove the validity of the method, it outperforms BoVW on image annotation precision ranged from 17.8% to 25.7%.

Key words: image annotation, Bag of Visual Word(BoVW) model, multiple scale space, multiple scale visual word, multiple kernel learning, weight optimization

中图分类号: 

• TP391