弱标签环境下基于语义邻域学习的图像标注

Image Annotation by Semantic Neighborhood Learning from Weakly Labeled Dataset

图像语义自动标注是实现图像语义检索与管理的关键,是具有挑战性的研究课题.传统的图像标注方法需要具有完整、准确标签的数据集才能取得较好的标注性能.然而,在现实应用中获得数据的标签往往是不准确、不完整的,并且标签分布不均衡.对于Web图像和社会化图像尤其如此.为了更好地利用这些弱标签样本,提出了一种基于语义邻域学习的图像自动标注方法(semantic neighborhood learning from weakly labeled image,SNLWL).首先在邻域标签损失误差最小化意义下,填充训练集样本标签.通过递进式的邻域选择过程,保证建立的语义一致邻域内样本具有全局相似性、部分相关性和语义一致性,并且语义标签分布平衡.在邻域标签重构误差最小化意义下进行标签预测,降低噪声标签对性能的影响.多个数据集上的实验结果表明,与已知的具有较好标注效果的方法相比,此方法更适用于处理弱标签数据集,标准评测集上的测试也表明了此方法的有效性.

With the advance of Web technology, image sharing has become much easier than ever before. Automatic image annotation, which can predict relevant labels for images, is becoming more and more important. Traditional image annotation methods usually require a large number of complete, accurate labeled data to obtain good annotation performance. However, since obtaining weak labeled training data is often much easier and costs less efforts than obtaining a large amount of fully labeled training data, image labels are often incomplete and inaccurate in real world environment. In addition, different labels usually have large frequency differences. To effectively harness these weakly labeled images, in this paper, an automatic image annotation approach based on semantic neighborhood learning （SNLWL） is proposed. The missing labels are replenished by minimizing the reweighted error functions on training data. Then, the semantic neighborhood is obtained by a progressive neighborhood construction approach. We incorporate label completeness, global similarity, conceptual similarity, and partly correlation into the stage. In addition, an effective label inference strategy is proposed by minimizing the neighborhood reconstruction error to handle the noise in the labels. Extensive experimental results on different benchmark datasets show that the proposed approach makes a marked improvement as compared with other methods.