NDT Method for Weld Defects Based on FMPVit Transformer Model

The primary NDT method for welding defects is the image-based detection. Currently, best performance detection based on transformer model. However, with its high accuracy, it has many limitations, such as large model parameters, data sample requirements, and expensive computer resources. This a weaker ability to capture local features compared global features. In this study, an improved optimized defect identification framework named Fast Multi-Path Vision (FMPVit) proposed uses multilayer parallel architecture enhances information of through advanced multiscale convolution feature aggregation addition new module. Finally, validation test carried out using open dataset weld seams. proven exhibit evident improvement over mainstream baseline.