Weld-FHG-YOLO: A Lightweight Multi-Frequency Feature Fusion Network for Weld Keypoint Localization

Abstract

Accurate weld keypoint localization is an important visual perception task for robotic welding, weld tracking, and intelligent manufacturing. However, weld keypoints in line-structured light images are usually small, weakly textured, and easily affected by reflections, noise, and spurious laser stripes. These factors make it difficult for lightweight detection models to balance localization accuracy and computational efficiency. To address this problem, this paper proposes Weld-FHG-YOLO, a lightweight multi-frequency feature fusion network for weld keypoint localization. The proposed model is built on the You Only Look Once version 11 nano (YOLO11n) framework and focuses on optimizing feature fusion and scale transformation in the Neck. Specifically, FasterC3K2 is introduced to replace the original C3K2 modules in the Neck, thereby reducing redundant computation during multi-scale feature fusion. In addition, a Haar Wavelet Decomposition and Group Shuffle Convolution (HWD-GSConv) downsampling fusion module is designed, in which Haar wavelet decomposition preserves low-frequency structural information and high-frequency details, while GSConv performs lightweight multi-frequency feature fusion. Experimental results show that Weld-FHG-YOLO achieves 2.301 M parameters and 6.016 GFLOPs, which are 11.26% and 6.83% lower than those of YOLO11n, respectively. Meanwhile, mAP@0.5:0.95 increases from 0.7337 to 0.7901, the Mean Center Error (MCE) decreases from 2.254 px to 2.131 px, and the CPU inference speed increases from 13.69 to 14.77 frames per second (FPS). These results indicate that the proposed method improves strict localization accuracy and localization stability while maintaining a lightweight computational profile, providing a practical visual perception approach for weld keypoint localization in resource-constrained intelligent manufacturing scenarios.