Lightweight DETR algorithm for X-ray weld defect detection.

Area of Science:

• Nondestructive Testing
• Artificial Intelligence
• Computer Vision

Background:

• Traditional target detection models struggle with low-contrast, multi-scale, and complex weld defects in X-ray images, especially under resource constraints.
• Computational limitations hinder the deployment of advanced defect detection systems in industrial settings.

Purpose of the Study:

• To propose a Lightweight Multi-Scale Real-Time non-convolution detection TRansformer (LMS-RTDETR) for efficient and accurate X-ray weld defect detection.
• To address the challenges of limited computational resources and storage space in industrial nondestructive testing.

Main Methods:

• Implemented a Multi-Scale Feature Aggregation (MSFA) module for parallel convolution and feature reorganization.
• Introduced an Efficient Additive Attention Feature Interaction module to reduce computational complexity while preserving contextual information.
• Utilized a Multi-Scale Feature Pyramid Network (MSFPN) for effective multi-branch feature fusion.
• Optimized bounding box regression with the NWD-Inner GIoU loss function.

Main Results:

• LMS-RTDETR reduced model parameters by 38.34% and floating-point operations by 29.30% on an X-ray weld defect dataset.
• Achieved a 3.10% improvement in mean Average Precision (mAP50:95).
• Demonstrated superior performance in detecting challenging weld defects.

Conclusions:

• LMS-RTDETR offers a high-precision, computationally efficient solution for X-ray weld defect detection.
• The proposed model is well-suited for resource-limited environments in industrial nondestructive testing.
• This work advances the application of deep learning in critical infrastructure inspection.