Abstract: Change detection in buildings holds significant importance in the fields of remote sensing image processing and pattern recognition. However, data annotation has always been a prominent challenge in the application of deep learning algorithms, especially in change detection scenarios. To address the data annotation challenges in deep learning-based change detection algorithms, this paper proposes an innovative semi-supervised learning method. This method employs a Siamese network that fuses bi-temporal features for feature extraction and constructs a teacher-student network framework for semi-supervised model training. To further enhance the accuracy of semi-supervised change detection, this paper introduces random perturbations in deep features to achieve consistency regularization. Additionally, on the level of image deep features, the paper proposes a method for forming decision boundaries by capturing differences in bi-temporal image features to distinguish changes in bi-temporal images. This method achieved Intersection over Union (IoU) scores of 83.04% and 85.57% on the Levir-CD and WHU Building datasets, respectively. Experimental results show that the proposed method can achieve performance levels close to fully supervised training with a limited amount of labeled data.