Scene text detection aims to accurately localize text instances in natural scenes. With impressive capability of deep learning methods, the related knowledge-based computer vision systems have achieved remarkable detection results in real-world scenarios. However, they generally require a large amount of training samples with huge manually labeling cost to extract sufficient vision knowledge. To address this problem, we propose a teacherstudent network for semi-supervised scene text detection. Specifically, we enhance which not only enhances knowledge learned by teacher network with pseudo-labeling enlarged training dataset, but also , which generates pseudo-labels through the teacher network and enlarge training dataset with pseudo-labeled samples, and then trains student network for detection with the enhanced dataset. The proposed method not only involves a pseudo-label quality assessment mechanism to improve the robustness of teacher network, but also designs a cascade hybrid framework in student network for informative information refinement. Experimental results demonstrate that our method achieves state-of-the-art performance on horizontal datasets (ICDAR2013), multi-oriented datasets (ICDAR2015), multilingual datasets (ICDAR2017-MLT) and arbitrary shape datasets (CTW1500, Total-Text).