Abstract: The potential image semantic information in feedback logs helps to improve the quality of image automatic annotation, but the noisy and one-sidedness of the log restrict its function. So it proposes a log-based collaborative automatic image annotation algorithm. It uses incremental association rule to mine log messages according to its characteristics to remove the noisy data and expand the amount of image annotation words with the thought of collaborative filtering. Combining the relationship of the annotation words from the WordNet and the low-level image features, it can accomplish image automatic annotation with adoption of hybrid probabilistic model. The experimental results based on Corel5K and Internet image set show that the proposed algorithm can reduce the negative effects from noisy and one-sidedness of the log data, and improve the performance and efficiency of the automatic image annotation.

Key words: collaborative filtering, hybrid probabilistic model, image automatic annotation, incremental association rules, log

摘要： 反馈日志隐含的图像语义信息有助于图像自动标注，但日志数据中存在的噪声、片面性等问题制约了其作用，故提出基于日志的协同图像自动标注算法。根据日志获取的特点，采用增量关联规则挖掘处理日志信息去除其噪声，利用协同滤波思想扩展图像标注词数量，利用WordNet得到标注词间关系，并结合图像底层特征利用混合概率模型实现图像自动标注。在Corel5K和互联网数据集上的实验表明：该算法降低了日志噪声及片面性所带来的影响，提高了图像自动标注效率和质量。

关键词: 协同滤波, 混合概率模型, 图像自动标注, 增量关联规则, 日志