Semisupervised semantic segmentation based on optimized pseudo-label

The pseudo-tags generated by the multi-teacher method in the semi-supervised semantic segmentation task have better accuracy than that of the single teacher model. The method usually averages the prediction results of multiple teacher models and uses them as pseudo-tags to learn the student model, so as to alleviate the instability of the single teacher model and reduce the coupling between the student model and the teacher model. However, multiple teacher models cannot obtain a result with high confidence for some pixel points. Difficult and simple pixels are separated by the result overlap screening judgment on the false labels generated by the teacher models. For multiple pixels with different categories predicted by teachers, the false labels generated will not contain positive categories, and only the undisputed categories will be used as negative samples for students to learn. So that the student network can not be disturbed by false labels with low confidence, and effectively reduce the impact of false label wrong prediction. Besides. Combined with the multi-stage learning method, by labeling unsupervised data several times, the advantage of more stable false labels generated by multiple teachers can be more fully utilized, thus further improving the accuracy of false labels. Experimental results on the extended dataset of PASCAL VOC 2012 show that networks using this method can achieve higher mIoU scores than other mainstream semi-supervised semantic segmentation models.