基于特征融合和代价敏感学习的图像标注方法

Image annotation method based on feature fusion and cost-sensitive learning

针对图像标注数据集中存在的标注对象比例不一致和标签分布不平衡问题,提出基于特征融合和代价敏感学习的图像标注方法。在卷积神经网络中加入特征融合层,改进VGG16原有的网络结构,特征融合层结合注意力机制,对网络中不同卷积层提取的多尺度特征进行选择性融合,提升对不同尺度对象的标注精度;将代价敏感学习融入损失函数对网络模型进行训练,提升网络的泛化性能。实验结果表明,该方法能提升图像标注的准确率,增加对低频标签的召回率。

To solve the problems of object scale inconsistency and category imbalance in image datasets,an image annotation method based on feature fusion and cost-sensitive learning was proposed.The feature fusion layer was added to the convolutional neural network to improve the original network structure of VGG16,and the attention mechanism was combined to selectively fuse the multi-scale features extracted from different convolutional layers in the network to improve the performance of objects of different scales.Cost-sensitive learning was incorporated into the loss function to train the network model to improve the genera-lization performance of the network.Experimental results show that the proposed method can improve the accuracy of image annotation and increase the recall rate of low-frequency labels.