Abstract: The performance of text-based image retrieval is highly dependent on manual tagging, and the deep learning can be used to realize image keywords generated automatically. Combining the predictions of many different large neural nets is an effective way for improving the classification accuracy. Firstly, for improving the generalization performance of the deep learning model, this paper proposes the Dropout algorithm. Dropout is a technique for addressing this problem by randomly dropping units (along with their connections) from the neural network during training. So the algorithm is equivalent to train many neural networks for prediction. Next, by the reason of the diverse keywords of image, this paper proposes a two steps algorithm for image annotation. First step, the keywords are divided into three parts: base keywords, candidate keywords and irrelevant keywords depending on the output of all neural networks. Second step, the keywords are chosen in candidate set depending on their correlation with base keywords. At last, the base keywords and chosen keywords are labeled for images. Conducting extensive experiments on three popular data sets, the results demonstrate that the proposed framework can achieve favorable performance for image annotation.

Key words:  image auto-annotation, deep learning, assemble learning, machine learning

摘要： 基于文本的图像检索技术强烈依赖于图像标签，深度学习可以用来实现图像标签的自动生成。多分类器融合是一种有效提升分类器精度的方法。为了提升深度学习模型的泛化性能，提出了Dropout算法。该方法的本质是在训练过程中随机地丢弃若干神经元，等价于同时训练多个子网络。由于图像标签的多样性，提出了两步标签融合算法：第一步，根据多个不同网络的输出将图像标签词汇分为基准词汇、备选词汇和无关词汇；第二步，选出备选词汇中与基准词汇强相关的词汇，基准词汇和被选出的词汇可作为图像的标签。最后，算法选取3个常用的数据集对提出的算法模型进行验证，实验结果表明，多分类器融合算法可以有效地解决图像自动标注问题。

关键词: 图像自动标注, 深度学习, 集成学习, 机器学习