Abstract: Background subtraction is often adopted in automatic weld defects inspection techniques based on real-time X-ray images. The key to this method is estimating of weld background, and it is mainly realized by low-pass filter or curve fitting methods. Generally, there are many false alarms and inaccurate defect shape due to inaccuracy of weld background obtained by these methods. In this paper a new background subtraction based on partial surface reconstruction was proposed. Firstly, the possible defect positions were identified by gray-level wave analysis, and then weld background was estimated by reconstruction in these areas. Results showed that false alarm and missed detection had been decreased effectively, and the defect shape had been more accurate.