SEMI-SUPERVISED SEMANTIC SEGMENTATION METHOD BASED ON ADVERSARIAL LEARNING AND CONSISTENT REGULARIZATION

In order to reduce the need of pixel level label in semantic segmentation task, this paper proposes a semi-supervised semantic segmentation method based on adversarial learning and Mean teachers model. The training process of this method was divided into two stages. In the first stage, the discriminating network was connected after the segmented network, and the prediction results of the segmentation network were gradually closer to the true label through adversarial learning. In the second stage, the network parameters of the first stage were used to do exponential moving average to get the teacher network to train the consistency with the segmented network, so that the performance of the model was further improved. Experiments on Pascal VOC 2012 data set show that under the same number of label training, the proposed method outperforms the existing semi-supervised semantic segmentation in quality of segmentation graph and evaluating indicator mIoU.