The extraction of bubbles from ice micrographs using deep learning methods requires a significant amount of annotated data. However， the manual annotation of bubbles presents a significant challenge in this regard. A domain-adaptive extraction method is proposed， which utilizes the CycleGAN style transfer network and the Attention U-Net image segmentation network. In this method， the image generated by simulating the shape of the bubble is used as the source domain， and the icing microscopic image is used as the target domain. The source domain image is converted into the target domain style through CycleGAN， and the Attention U-Net network is trained using the style-converted source domain dataset. The two cases of unlabeled icing images and a few labeled images are verified by comparative experiments. Experimental results show that the unsupervised extraction of air bubbles from icing microscopic images can be achieved without annotated images， and the method can achieve more accurate air bubble extraction with only a few annotated images.