基于多教师网络模型的半监督语义分割方法

Semi-supervised Semantic Segmentation Method Based on Multiple Teacher Network Model

基于一致性正则化的方法在半监督语义分割任务中展现出了较好的性能,这类方法通常涉及两个角色:一个显式或隐式的教师网络和一个学生网络。其中学生网络通过最小化两个网络对不同扰动样本预测结果之间的一致性损失实现训练。但是来自单个教师网络的不可靠预测可能会导致学生网络学习到错误的信息。通过将平均教师模型MT的单教师网络扩展为多教师网络,提出了多平均教师网络(Multiple Mean Teacher Network,MMTNet)模型,使学生网络从多个教师网络的平均预测结果进行学习,有效降低单个教师网络预测错误的影响。此外,MMTNet通过对无标签数据进行强、弱数据增强的方式对无标签数据进行数据扰动,增加了无标签数据的多样性,在一定程度上缓解了学生网络和教师网络之间存在的耦合问题,避免了学生网络对教师网络的过度拟合,从而进一步降低了教师网络进行伪标签预测错误时所产生的影响。在PASCAL VOC 2012扩充数据集上的实验结果表明,所提出的多平均教师网络MMTNet模型可获得比其他目前主流的半监督语义分割方法更高的平均交并比,且实际分割效果更优。

The methods based on consistency regularization show better performance in semi-supervised semantic segmentation task.Such methods usually involve two roles,an explicit or implicit teacher network,and a student network which is trained by minimizing the consistency loss between the prediction results of two networks for different perturbation samples.But unreliable predictions from a single-teacher network may cause the student network to learn wrong information.By extending the mean teacher(MT)model to the multiple teacher network,multiplemeanteacher network(MMTNet)is proposed to make the student network learn from the average prediction results of multiple teacher networks,which can effectively reduce the impact of single-teacher network prediction errors.In addition,MMTNet implements data perturbation of unlabeled data by applying strong data augmentation and weak data augmentation to the unlabeled data,which increases the diversity of unlabeled data,alleviates the coupling problem between student network and teacher network to a certain extent and avoids the overfitting of student network to teacher network,so as to further reduce the impact of pseudo-label prediction errors in the teacher network.Experimental results on VOC 2012 augmented dataset show that the proposed multiple mean teacher network model MMTNet can achieve higher mean intersection over union than other mainstream semi-supervised semantic segmentation methods,and the actual segmentation performance is better.