Weld defect detection is an important research topic in the field of industrial non-destructive testing. However, this is a challenging task, as X-ray images typically exhibit low contrast and defects often have varying shapes and sizes, making existing methods unable to accurately capture the location information of weld defects. To address these challenges, this paper develops a new framework to effective detect different types of defects from low quality X-ray images. Firstly, an adaptive contrast enhancement method is designed to effectively generate optimized X-ray images, which is beneficial for the feature extraction process. Secondly, an adaptive feature pyramid network equipped with deformable convolution is proposed to fit defects with varying shapes and sizes, effectively improving the generalization performance of the model. In practical applications, we adopt the pipeline weld X-ray defect dataset in northern China and demonstrate the effectiveness of the method.