YOLO-Weld: A Modified YOLOv5-Based Weld Feature Detection Network for Extreme Weld Noise

Abstract

Weld feature point detection is a key technology for welding trajectory planning and tracking. Existing two-stage detection methods and conventional convolutional neural network (CNN)-based approaches encounter performance bottlenecks under extreme welding noise conditions. To better obtain accurate weld feature point locations in high-noise environments, we propose a feature point detection network, YOLO-Weld, based on an improved You Only Look Once version 5 (YOLOv5). By introducing the reparameterized convolutional neural network (RepVGG) module, the network structure is optimized, enhancing detection speed. The utilization of a normalization-based attention module (NAM) in the network enhances the network’s perception of feature points. A lightweight decoupled head, RD-Head, is designed to improve classification and regression accuracy. Furthermore, a welding noise generation method is proposed, increasing the model’s robustness in extreme noise environments. Finally, the model is tested on a custom dataset of five weld types, demonstrating better performance than two-stage detection methods and conventional CNN approaches. The proposed model can accurately detect feature points in high-noise environments while meeting real-time welding requirements. In terms of the model’s performance, the average error of detecting feature points in images is 2.100 pixels, while the average error in the world coordinate system is 0.114 mm, sufficiently meeting the accuracy needs of various practical welding tasks.

High-accuracy and lightweight weld surface defect detector based on graph convolution decoupling head

“…Then, when the streamlined architecture is applied in method 2, the missing detection rates of both weld and defects show a significant increase. Such a result is different from that of [20], which means that the use of DWconv in the unoptimized feature extraction part will interfere with the detection and make the target unrecognizable. This is attributed to that the separable convolution may lose some important information and cannot make up for it, resulting in multiple ambiguous samples appearing.…”

“…Model size and detection speed also hold great significance. Complex feature extraction structures or strategies, when newly introduced, may prove incompatible with the original network, thereby increasing network complexity and prolonging inference times [20].…”

“…While retaining the efficient feature extraction backbone and neck parts, the model's main improvement focuses on the detection decoupled head and label assignment strategy to mitigate information conflicts and improve detection accuracy. Although the current double decoupled head can alleviate the conflicts and redundant information to improve accuracy, it will lead to a great increase in model complexity [20]. Notably, the emerging graph convolutional network (GCN) [21] can improve the ability to read and recognize surface defect features, as it can compensate for the limitations of CNN in dealing with graph structures in Euclidean space [22].…”