Currently, crack detection techniques based on semantic segmentation heavily rely on a large number of manually annotated high-precision pixel-level labels. However, the challenging, time-consuming, unstable manual annotation hinders the acquisition of a large quantity of high-precision pixel-level labels. In this paper, we introduce semi-supervised techniques to lessen the dependency on high-precision data and improve existing semi-supervised techniques for crack segmentation scenarios. Firstly, considering the high complexity and variability of cracks, which pose challenges in terms of model generalization, we draw inspiration from multi-task learning and extend a separate student model based on the conventional teacher-student framework. The two student models learn from labelled and unlabelled crack images, respectively, and interact through a shared encoder, thus improving the model's generalization capability. Secondly, considering the significant sample imbalance and feature disparity in crack data, we introduce a GEC-Loss to prevent the model from merely leaning simple crack features while ignoring complex and crucial ones. Additionally, to enhance the consistency capacity of the semi-supervised model, we propose a new data augmentation method called Affine-Paste. This method expands multiple geometric perspectives for crack detection and improves viewpoint consistency in the semi-supervised semantic segmentation task. Extensive experiments are conducted on our self-made dataset and public datasets. The far better performance over other recent semi-supervised methods demonstrates the effectiveness and superiority of our proposed method.