Aiming at the problem of low defect recognition rate in radiographic images, background estimation and differential operations are used to enhance defects and suppress complex background and noise. The method first used the mask image obtained by Otsu segmentation to extract the weld area; Secondly, the background estimation of the weld area was performed by the improved median filter, and the difference image containing the defect was obtained by inverting the background difference; Then, according to the difference of the gradient direction at the edge of the defect and the false detection, the multi-directional and multi-level gradient was used to effectively solve the background residual problem; Finally, t the differential image containing the defect was binarized by adaptive threshold segmentation. After experimental simulation, this method has a high defect recognition rate, with a recall rate and an accuracy rate of 91.90% and 90.95%, respectively, which has good application value in practice.