In order to effectively detect crack, a crack segmentation method using multi-scale structured forests and wavelet transform is proposed to improve robust performance of crack detection. The multi-channel feature extraction of crack image, and discrete mapping of the corresponding ground truth is carried out respectively with assistance of multiple crack image and ground truth. Triangle filter and down-sample are adopted to process regularity candidate features and correlation candidate features, which are used to train and validate structured forest classifier. And, structured forest classifier is used to crack segmentation of test images in multi-scale. According to experiment results in 776 structural crack image and 600 steel beam image datasets, the proposed method can obtain highest segmentation accuracy in a short time than single multi-scale structured forest method and other segmentation methods.