Improved Lightweight Rebar Detection Network Based on YOLOv8s Algorithm
Zhichao Hu, Yi Wang, WU Jun-ping, Xiong Wanli, Li Bilian
Abstract
This paper introduces an improved YOLOv8-based algorithm, Yolo-Rebar, designed to address the challenges of intelligent rebar counting in construction engineering. By integrating SPDConv to replace traditional convolution for downsampling, and combining bi-directional feature pyramid networks (Bi-PAN-FPN), internal intersection over union (Inner-IoU) evaluation strategy, and Dynamic Head component, Yolo-Rebar optimizes the network structure and inference process, significantly reducing computational load and parameter count while maintaining a high detection accuracy (mAP of 0.985). Maintaining a low computational demand (29.9 GFLOPs) and a moderate model size (27.1 MB), Yolo-Rebar outperforms Yolov5s and Yolov8s models in detection accuracy by 1.5% and 0.8% respectively, and compared to Yolov3-app and Yolov8m models, it requires lower computational resources while maintaining high accuracy. Empirical results demonstrate that Yolo-Rebar exhibits remarkable robustness and precision in complex construction environments, such as varying lighting conditions, rebar stacking, and occlusions. This research not only enhances the efficiency and accuracy of material acceptance in construction engineering but also provides a new direction for the further development of deep learning technology in industrial applications.