Abstract: Concerning the imbalance of the data set used in image annotation, a new self-balancing model based on external database was proposed. Firstly, the low-frequency points were found based on word frequency distribution of the original database, and an appropriate amount of image was added from an external database under the self-balancing mode for each low-frequency word. Secondly, the image features were extracted, and 47065 visual vocabulary of the original data set and 996 visual words extracted from additional images of external databases were clustered together. Lastly, each image was annotated by the improved image automatic annotation model based on external database. The proposed method overcomes the imbalance in image annotation, making the number of words which can be correctly labeled at least once, precision and recall be increased obviously.

Key words: external database, self-balancing mode, database imbalance, image automatic annotation

摘要： 针对图像标注中所使用数据集存在的数据不平衡问题，提出一种新的基于外部数据库的自动平衡模型。该模型先依据原始数据库中词频分布来找出低频点，再根据自动平衡模式，对每个低频词，从外部数据库中增加相应的图片;然后对图片进行特征提取，对Corel 5k数据集中的47065个视觉词汇和从外部数据库中追加的图片中提取出来的996个视觉词汇进行聚类;最后利用基于外部数据库的图像自动标注改善模型对图像进行标注。此方法克服了图像标注中数据库存在的不平衡问题，使得至少被正确标注一次的词的数量、精确率和召回率等均明显提高。

关键词: 外部数据库, 自动平衡模式, 数据库不平衡, 图像自动标注