Pseudo-training Method For Semi-supervised Semantic Segmentation

In recent years,to alleviate the shortage of pixel-level annotation samples,the self-training based semi-supervised semantic segmentation method has attracted wide attention.To expand the number of training samples in the data set,these methods first train the segmentation network in a small number of labeled samples and then take the prediction results of a large number of unlabeled samples as false labels.However,the performance of the model is limited because there is a lot of noise in the pseudo label.In addition,when the number of labeled samples is much smaller than the number of pseudo-labeled,the noise within the pseudo label will degrade the performance of the model.To effectively alleviate the negative impact of noise in pseudo-label,in this paper,we deeply study the pseudo-training based semi-supervised semantic segmentation method.The main contributions are as follows:Firstly,we analyze three problems existing in the pseudo-training strategy.And then we study and propose a cooperative pseudo-training based semi-supervised image semantic segmentation model,which includes a high-level semantic error correction module,a low-level semantic error correction module,and a dual-student based segmentation network.In addition,we proposed an adaptive weighted cross entropy loss function.The goal of the high-level semantic error correction module and the low-level semantic error correction module is to alleviate the false positive and false negative classification in pseudo-labels.The purpose of the dual-student segmentation framework is to overcome the shortcoming of the single segmentation based pseudo-training method.The adaptive weighted cross entropy loss function is proposed to solve the problem of pseudo-label waste caused by the confidence degree based threshold filtering method.Secondly,we analyze the potential negative optimization problem of the single branch based semantic segmentation model in pseudo-training method.And then we study and proposed a dual branch pseudo-training based semi-supervised image semantic segmentation model,which includes a dual-branch segmentation network and an improved single-branch decoder.The proposed dual-branch segmentation network consists of a shared encoder and two independent decoder branches to reduce the interference of weak supervised signals on model training.An improved DeepLabv3+ decoder module is proposed to improve the effectiveness of the feature fusion.Finally,we conduct a large number of ablation experiments and comparison experiments on the proposed two pseudo-training based semi-supervised image semantic segmentation methods.Experimental results verify the effectiveness of the proposed two pseudo-training based semi-supervised models.And the experimental results show that the proposed method is superior to other advanced semi-supervised semantic segmentation methods.