Leveraging Convolutional Neural Networks for Weld Defect Detection: An Innovative SDR Approach

This paper proposes a method for detecting welding defects in oil and gas pipeline welds by combining Convolutional Neural Networks (CNNs) with Suspected Defect Regions (SDRs) for non-destructive testing. X-ray images of welds are often compromised by high levels of noise and low contrast. To mitigate these issues, we apply mean filtering and sine function enhancement to process the images. We then introduce a gray density-based SDR segmentation technique and employ CNNs for defect identification within the segmented SDRs. Experimental results show that the proposed method significantly enhances the automation of welding defect detection, achieving an accuracy of 98.9%.