YOLOv8n-BAseg Instance Segmentation Network For Angle Tower Bolt Groups and Connecting Plates Detection

The fastening of angle tower bolts is important in tower solidity. And angle tower bolt detection is an important part of the angle tower bolt fastening task. The first step of bolt detection is to detect the angle tower bolt groups and tower connecting plates, which is convenient for the subsequent detection of angle tower bolts and planning of the robot arm movement path. Due to the irregular labeling of the angle tower bolt groups and tower connecting plates data, an instance segmentation network is used for detection. In this paper, based on YOLOv8-seg instance segmentation network, the YOLOv8-BAseg instance segmentation network is proposed by adopting the bidirectional cross-scale connection and weighted feature fusion mechanism BiFPN_Add2 instead of the PAFPN feature fusion mechanism. The experimental results show that the YOLOv8-BAseg network of size n has fewer parameters, lower model complexity, providing the possibility of real-time detection when deployed on mobile devices. The YOLOv8n-BAseg network is trained on the Angle Tower Bolt image dataset, and the mAP50 reaches 95.3%, and the speed reaches 2.5 ms per image. Compared with the yolov8n-seg instance segmentation network, its segmentation accuracy is improved by 0.4%, and 2 ms improves the segmentation speed of each image. The results of the comparative experiments show that the improved model performs well in terms of model complexity, inference speed, and segmentation accuracy, which are significantly better than the existing instance segmentation model. In conclusion, the YOLOv8n-BAseg model balances the model performance and computational complexity and better meets the needs of angle steel tower bolt groups and connecting plates detection.