To address the detection challenges in X-ray weld seam images caused by weak contrast, slender structures, and multi-scale coexistence, we propose MGDR-YOLO, an industrially deployable detector with four coordinated designs. First, a MultiBackbone parallel heterogeneous backbone is designed to perform complementary direction–detail modeling and lightweight context modeling under a shared shallow stem, enhancing the joint representation of fine-grained features and global semantics. Second, Gated Attention Fusion Block (GAFB) is introduced to perform selective in-scale fusion via channel gating and local–global attention mechanisms, thereby suppressing channel redundancy and noise leakage induced by naive concatenation. Third, Directional Feature Convolution (DFConv) decouples standard 2D convolution into horizontal and vertical branches and fuses them using depthwise separable convolution, substantially reducing computational cost while preserving geometric alignment. Finally, Rep Shared Convolutional Detection Head (RSCD) improves detection head consistency and inference throughput through cross-scale shared convolutions and a training-to-deployment re-parameterization scheme. The experimental results show that MGDR-YOLO significantly outperforms YOLOv11n, increasing the mean average precision (mAP) from 92.9% to 95.2%. The performance gain is most pronounced for the LP class (slender and low-contrast defects), with an mAP improvement of 10.1 percentage points. Meanwhile, the proposed model achieves a 39.4% increase in frames per second (FPS) while reducing the number of parameters by 46.2%, demonstrating superior efficiency. These results indicate that MGDR-YOLO consistently improves the accuracy and robustness of X-ray weld defect detection while maintaining real-time performance, making it well suited for resource-constrained industrial online inspection scenarios.
1 March 2026 in IAES International Journal of Robotics and Automation (IJRA)
To address challenges like global drift, unstable matching, and high computational cost in light detection and ranging simultaneous localization and mapping (LiDAR SLAM) under complex conditions, this paper proposes an improved algorithm based on the LeGO-LOAM framework. A Newton-optimized normal distributions transform (NDT) is integrated to improve point cloud registration by constructing a negative log-likelihood objective and optimizing pose estimation. Using initial pose information from LeGO-LOAM accelerates convergence and enhances system robustness. This work addresses the problem of insufficient adaptability of existing algorithms in real scenarios. By deploying an independently designed four-wheel omnidirectional mobile robot platform, a hybrid LiDAR SLAM framework is used for precise positioning and map construction in complex campus environments, successfully reducing the positioning error to the centimeter level. Experiments on the KITTI dataset show a 43.51% reduction in maximum localization error and a 30.83% decrease in average error. Field tests in real-world campus environments with pedestrians, bicycles, and vehicles demonstrate strong reliability, adaptability, and resistance to interference. Horizontal error was reduced by about 58.26%, lowering the average error from 4.60 m to 1.92 m. Although computational load increases, it is offset by using high-performance LiDAR and processors. The enhanced accuracy and drift reduction significantly outperform traditional methods. At critical time points such as 50 seconds and 100 seconds, the system achieved high-precision pose estimation and accurate environmental reconstruction.