Welding defect classification method for pipelines using improved cascade attention mechanism and edge detection

Accurate welding surface defect detection is key to maintaining pipeline health and service. These defects are often mixed and cause serious background noise. In addition, traditional methods are limited to small sample datasets. This study presents a novel approach for the classification of pipeline welding defects. First, an attention-enhanced deep convolutional GAN that integrates a self-attention module is designed to enlarge the training dataset. Second, to remove noise from the image edge detection process, an edge detection method integrating the Canny operator and Modified Mathematical Morphology is introduced. In this method, the Canny operator’s double-threshold filtering is omitted; instead, morphological erosion removes short edges, and morphological dilation reconnects discontinuous edges. Finally, the multi-cascade attention mechanism is embedded into the VGG16 network as a classification model, where spatial attention is applied to low-level features to improve microscopic defects detection accuracy, while channel attention is employed on high-level features to enhance the integrity of macroscopic defect detection. Method validation is conducted on both self-developed and public datasets. The experimental results show that the proposed method achieved an average classification accuracy of 99.05% on the self-developed dataset and a precision of 99.18% on the public dataset, indicating superior stability in classification.