The crack-detection method based on deep learning relies heavily on a large amount of pixel-level annotation information. Thus, a crack-detection method based on semi-supervized learning is proposed. The proposed method introduces multiscale modules into the network model of crack-detection. It uses only a small part of pixel-level annotation data for fully-supervized training. For the unlabeled data, the fusion of multiple saliency area detection methods to generate pseudolabels can reduce pixel-level reliance on the labeled information. The improved network is experimentally verified on the crack dataset. It is compared with the commonly used semantic segmentation network and weakly supervized experimental benchmarks from the perspective of subjective evaluation, accuracy, recall rate, and F1-score. The experimental results show that the improved network can effectively enhance crack recognition accuracy. The proposed semi-supervized training method can achieve recognition accuracy and recall rate equivalent to the fully-supervized method when only 6.25% pixel-level label information is required.