Segmentation of X-ray images of teeth plays an important role in dental diagnosis. However, the labeling information is expensive, and it is of great significance to effectively use the unlabeled images to improve the segmentation performance. In order to make full use of unlabeled data, in this paper, we design a three-stage pseudo-label training framework based on self-training to improve the pseudo-label quality in a progressive way. Hard data augmentation is introduced to act on the pseudo-labels to mitigate the model overfitting to the noise in the pseudo-labels. In addition, we designed segmentation model with denoiser to improve the segmentation result while reducing the noise in pseudo-labels by removing the noise from the segmentation map generated by the segmentation model. Our method achieved an Dice score of 93.15%, an IoU score of 98.03% and HD score of 97.68% for the teeth segmentation on the validation set using a NVIDIA GeForce RTX 3090. The average training time was 2 h.