【摘要】 针对数据集标注成本高昂的问题，提出一种半监督语义分割算法，可以利用少量标注数据和大量无标注数据训练高精度的分割模型，节省大量人工标注成本。该文算法在经典的平均教师算法的基础上进行改进，将单组教师-学生模型扩展为双分支结构，采用交叉监督的方式进行学生模型训练，有效解决了平均教师算法的模型耦合问题。另外，引入非线性投影模块构造两个非对称分支，使模型学习到异质知识，进一步提升了算法性能。对算法进行针对性调整后，可以较好地适配变电站实际应用场景。实验结果表明，该文算法不仅在变电站表计图像分割任务中表现出良好的性能，还在两个学术数据集PASCAL VOC 2012和CamVid中超过了当前主流算法。更多还原

【Abstract】 Aiming at the problem of expensive dataset annotation costs, a semi-supervised semantic segmentation algorithm is proposed. This algorithm is able to utilize a small set of labeled data and a large set of unlabeled data to train a high-performance segmentation model; therefore a lot of manual annotation costs can be saved. The proposed method improves the classic mean teacher algorithm by extending the single teacher-student branch to a dual-branch structure, where the student models are trained in a cross-supervision manner. In this way, the model coupling issue of the vanilla mean teacher algorithm is effectively solved. To further improve the performance, a nonlinear projector is introduced to make two branches asymmetric, which helps the model learn heterogeneous knowledge. With some minor modifications, the algorithm can well adapt to practical scenarios of substations.Experiment results demonstrate that the proposed method not only performs well on substation meter image segmentation task, but also achieves state-of-the-art performance on two academic datasets, i.e. PASCAL VOC 2012 and CamVid.更多还原