The crimping quality of the strain clamp in the transmission line affects the safety of the power grid. At present, the quality inspection method of the strain clamp crimping is mainly to take X-ray images and perform manual identification. However, due to the small size and tightly packed of the defect parts in the X-ray image of the strain clamp, the manual method appears to be time-consuming and labor-intensive and the accuracy rate is not high. Aiming at the above problems, an X-ray image detection system for crimping defects of strain clamps based on deep learning is proposed. The principle of hierarchical detection is adopted, Firstly, the CenterNet algorithm is used to locate the defective crimping part and cut out the crimping part to increase the proportion of crimping defects in the image. Secondly, the data are used to enhance the data set, and finally the RetinaNet algorithm is used to detect the crimp defect. By verification that the hierarchical detection strategy in this paper improves to a certain extent in accuracy and detection speed compared with the traditional detection algorithm, which meets the application requirements in actual engineering.