基于按类视觉词袋模型与标签投票的图像检索

摘要

视觉词袋模型在基于内容的图像检索中已经得到了广泛应用，然而对于自然图像的检索，由于图像目标视角多样、背景复杂、光照多变等原因，传统的视觉词袋模型的检索准确率仍然比较低。提出一种按类视觉词袋模型，即采用按照图像中目标物体的类别进行分组训练视觉词，从而提高视觉词的表征能力，改善检索效果；并根据检索返回图像的标签，以投票方式对查询目标做出判别，辅以标签检索，从而较大地提高了检索结果的准确率。%Bag of visual words model has been wildly adopted for content based image retrieval. However, regarding natural scene image retrieval, traditional bag of visual words model still bears relatively low retrieval precision, in the presence of various and complicated viewing angle, background, and illumination conditions. In this paper, a bag of categorized visual words model is pro-posed. In this model, words are obtained from categorized objects to improve its description ability and therefore improve retrieval accuracy. Thereafter tag voting is also employed to judge the query object from retrieved image, and tag retrieval is auxiliary to im-prove the retrieval performance.