Defects detection in weld joints based on visual attention and deep learning

Non-Destructive Testing (NDT) for pipeline weld joint defects is an important measure to ensure the safety of pipeline transportation. Manual inspection is the traditional detection method for the defects in the X-ray images, but it is easily affected by some factors such as the quality of the film and human physiological state, which will greatly decrease the accuracy and efficiency of defects detection. For different size of defects in weld joints, weak contrast and wide boundary transitions in the X-ray images, our model integrated Feature Pyramid Network FPN on the basis of Faster-RCNN was proposed and a new visual attention mechanism SPAM (Squeeze and Position Attention Mechanism) was introduced. And the data augmentation method based on geometry transformation is used to simulate defects at different places with different shapes. Experimental results show that the detection accuracy of the proposed model is better than Faster-RCNN, and the mAP value is increased by about 4.0%.