Abstract
Abstract
Short chopped glass fiber has attracted great attention due to its wide range of applications in the fields of construction, aviation, automotive, electronics, and others. To ensure consistent length in the glass fiber shortening process, a vision detection system based on the improved YOLOv8-OBB algorithm is proposed. The YOLOv8-OBB algorithm leverages the oriented bounding box (OBB) technology, which has enormous advantages in detecting targets with the features of dense distribution, disproportionate ratio of length to width, and also arrangement with arbitrary directions. Then, the YOLOv8-OBB algorithm can achieve a higher accuracy in object localization by outputting extra rotational angles, compared with the standard YOLOv8 based on traditional horizontal bounding boxes (HBB). Furthermore, in order to improve the object detection accuracy and also facilitate the deployment of the vision detection model on microprocessor systems, the YOLOv8-OBB algorithm has been improved and accelerated by substituting the original C2f module for C2f-Faster and C2f-Faster-EMA modules, which are more efficient and lightweight than the former. Additionally, a Global Context Block has been integrated into the Neck of the algorithm to enhance the global feature extraction capability of YOLOv8-OBB. Test results on the short-chopped glass fiber dataset show that the improved model achieves a notable 3.9% increase in mAP
50
compared to the baseline, while the GFLOPs and parameters are respectively reduced by 19.3% and 22.7%. Additionally, the comparisons with four rotational object detection models, i.e., Rotated-FCOS, Oriented-RCNN, Rotated-Faster-RCNN, and Rotated-RetinaNet, further provide an initial validation of the feasibility of our proposed method. Ultimately, the improved YOLOv8-OBB model was successfully deployed on the Jetson Nano to feasibly monitor the length of short chopped glass fiber in real time. The current system serves as a deployment-oriented proof of concept, providing a baseline for future industrial automated trend monitoring.