Currently, high-precision fully supervised methods are mainly used in the change detection tasks of high-resolution remote sensing images, which require a large amount of labeled data. However, labeling huge amounts of sample data is costly. In order to solve this limitation, a SemiPSENet based on consistent regularization is proposed. SemiPSENet consists of two parts: a supervised phase and an unsupervised phase, which is trained jointly with a small amount of labeled data and a large amount of unlabeled data. In the supervision phase, to generate a multiscale feature map that can fuse more context information, a GELU_PSP module is designed, which can greatly reduce the loss of semantic information. The squeeze-and-excitation (SE) attention module is added after GELU_PSP module, which can highlight important changing features and neglect irrelevant information based on the importance of each feature channel. In the unsupervised phase, to process the unlabeled data, the consistency regularization method is adopted, which can make the predictive change map have the consistency ability under different random perturbations, so that the semantic information can fully be used in the unlabeled image. Experiments on two public datasets LEVIR-CD and WHU show that the proposed SemiPSENet can achieve excellent detection results by relying on the training of a small amount of labeled data and a large amount of unlabeled data. Notably, our network performs better than other SOTA methods when using the same amount of labeled data.