Weld seam defect detection based on deformable convolutional neural networks

抄録

With the rapid advancement of smart manufacturing, automated detection of welding defects plays a crucial role in ensuring product quality and enhancing the efficiency and stability of production processes. Traditional manual detection methods struggle to meet modern production needs due to low accuracy, poor efficiency, and subjectivity. This paper utilizes a deformable convolutional neural network based on YOLOv5m to improve the accuracy of weld seam defect detection in X-ray images. It introduces deformable convolution kernels to identify irregular welding defects and employs decision and memory modules, proposing a feature repetition unit structure to optimize the network by reducing parameters and enhancing learning for small samples. Through comparative analysis with the original network, the improved deformable convolutional neural network shows significant improvements in loss, precision, and mAP metrics on small samples.