Detection of Weak Defects in Weld Joints Based on Poisson Fusion and Deep Learning

There exist crack, bar and round defects in the weld joints of pressure vessels and pipes, which are detected by the X radiographic inspection system. The check and evaluation for defects in the radiographic images are often done manually and the work efficiency is low, the evaluation has artificial subjectivity. The automatic detection method of defects in weld joints based on Poisson fusion and deep learning is proposed. The defects in weld joints are typical industrial weak objects because of their small size, distinct edge erosion, and small SNR in radiographic images. The training images should be augmented for the application of deep learning in the object detection of industrial images. The augmentation method based on Poisson image fusion is given to simulate the edge erosion in radiographic images. The histogram is used to find a suitable position for the Poisson fusion of the object defect and background region. The detection model is obtained in the framework of Faster R-CNN with the pre-trained ResNet50. Feature Pyramid Network is integrated for its strong detection capacity for industrial weak objects. The results show that our Poisson image fusion has a much greater contribution to the detection model than the general data augmentation in geometry transform in terms of AP and Recall.