A weld feature points detection method based on improved YOLO for welding robots in strong noise environment

The real-time recognition of weld feature locations in compounded noises is very critical when it comes to robotic weld tracking and intelligent welding. Due to strong noise interference, it is arduous for traditional machine vision to obtain satisfactory detection results. Traditional deep convolution networks carry a huge number of parameters including slow detection speed, and low detection accuracy, which cannot be reached the actual welding requirements. Therefore, this paper proposed a new lightweight detector called Light-YOLO-Welding based on an improved YOLOv4 for detecting weld feature points. Firstly, modification of the backbone network is done to be MobileNetv3, which decreases the network parameters and increases the detection speed. Secondly, the replacement of the path aggregation network is done with Bidirectional Feature Pyramid Network (BiFPN) based on the idea of bidirectional cross-scale connection that integrates richer semantic features and maintains spatial information, applying depthwise separable convolutions in SPP and BiFPN layers. Lastly, to solve the problem of imbalanced positive and negative samples in the assembled data, a class loss function based on AP-Loss is proposed to improve weld feature recognition accuracy. The results show that the mAP of this method is 98.84%, the accuracy is 98.08%, the recall is 97.87%, and the speed is 62.22f/s; the results further confirmed that the method has higher accuracy and speed. The average error of the camera coordinate system key weld points extraction is 0.102 mm. Compared with other methods, the proposed method is more reliable and efficient for detecting weld feature points.

Automated summary

This paper proposes a lightweight detector called Light-YOLO-Welding based on an improved YOLOv4 for real-time recognition of weld feature locations in compounded noises. By modifying the backbone network and using bidirectional feature pyramid network, the method achieves significant improvement in detection speed and accuracy. Experimental results show that this method has higher mAP and accuracy, and is more reliable and efficient compared to other methods.