基于迁移学习与权重支持向量机的图像多标签标注算法

为解决图像的多标签自动标注中标签不平衡性的问题,提出了一种基于迁移学习与权重支持向量机的图像自动标注方法。为了解决所选数据集规模较小无法训练出最优的卷积神经网络的问题,文中采用迁移学习的方法,将通过Imagenet数据集训练出的Alexnet的参数迁移到文中所用的卷积神经网络模型中,并对最后一层全连接层进行微调,利用多标签分类多合页损失函数构成多分类的支持向量机。最后,文中对低频标签进行权重排序以得到图像的多标签标注结果。在Corel-5k、Esp-Game和IAPR-TC12共3个数据集上进行了实验,权重支持向量机获得的平均召回率分别提升了10%、9%和6%,低频标签对其平均精确率均提升了12%。实验结果表明,基于迁移学习的权重支持向量机的图像多标签标注方法可在有效提高数据集的召回率的同时提升低频标签的平均精确度。

Multi-label Image Annotation Algorithm Based on Transfer Learning and Weighted Ranking Support Vector Machine

In order to resolve the class imbalance problem in multi-label image annotation, an improved annotation method based on transfer learning and WRSVM was proposed in this paper. As it was difficult to train a CNN from scratch by using small datasets, transfer learning was adopted to transfer the parameters of Alexnet trained by Imagenet dataset to the convolutional neural network model utilized in the study. Besides, the last fully connected layer was fine-tuned and the multi-label multi-hinge loss function was applied to constitute multi-class support vector machine. Finally, the weighted ranking was used to label the low-frequency labels to obtain the multi-label image annotation results. The experiments were performed on three datasets including Corel-5k, Esp-Game and IAPR-TC12. The experimental results showed that the average recall of the proposed method increased 10%, 9%, and 6%, respectively, and the average of precision increased 12% for the low-frequency labels, indicating the proposed CNN-WRSVM method could improve the average of recall and the average of precision for low-frequency labels.