Research on wire rod knot detection using an improved YOLOv8 algorithm

To address the challenge of quickly identifying knots in bundled wire rods, a real-time detection method for wire rod knots is proposed based on an improved YOLOv8 algorithm. By constructing and annotating a wire rod knot dataset, this method integrates a weighted Bi-directional Feature Pyramid Network (BIFPN) and Coordinate Attention (CA) mechanism into the YOLOv8 framework, enhancing detection effectiveness for knots of various sizes. Additionally, the application of EIoU Loss as the loss function further improves accuracy. Experimental results indicate the model rapidly and accurately detects knots of different sizes against complex backgrounds, achieving a 3.3% improvement in mAP 0.5 and a 3.6% increase in mAP 0.5:0.95. The model achieves a compact parameter size of 4.3MB, and its computational requirement is decreased to 7.1G, making it more lightweight in comparison with original YOLOv8. This reduction significantly saves training and detection time, which is well-suited for real-time detection tasks.