An Improved YOLOv8-Dyhead-WiseIoU Model for Positioning and Counting Detection of Grouting Sleeves in a Prefabricated Wall

Abstract

In China, despite the industrialization of prefabricated building production and construction, inspection is still mainly done manually, which has been criticized for subjectivity, fatigue, and high cost. Existing inspection methods rely on traditional visual models or dedicated equipment, which are susceptible to noise and errors of the equipment itself, causing them to lack robustness and lag behind production speed. In addition, there is a serious lack of automated inspection technology for the production stage. To address these challenges, this paper proposes a you only look once version 8 (YOLOv8)-based model that enhances the dynamic head (Dyhead) and wise intersection over union (WiseIOU) loss functions for accurate inspection of grout sleeves in prefabricated walls. A dedicated grout sleeve dataset is created to fill the data gap, and a compatible inspection scheme is designed for seamless integration into the production line. Using mean average precision (mAP) as the evaluation criterion, the developed YOLOv8-Dyhead-WiseIoU model is compared with traditional detection models such as Region-based Convolutional Neural Networks and Single Shot Multi-Box Detector, as well as other variants of YOLOv8. It has excellent performance in mean average precision of IOU thresholds of 50% (94.6%), 75% (86.7%), and 50%–95% (75.1%), and its volume is only 10.86 Megabytes (MB), which is 0.34 MB smaller than the basic model. Its compact size enables it to be quickly and cost-effectively deployed in portable detection equipment. The developed inspection scheme is deployed on the actual production line for detailed multicase tests. The results confirm the effectiveness of the proposed inspection method, filling the research gap in the current production process automation inspection, and providing a new way to improve the industrialization level of prefabricated parts quality inspection. This study develops a deep learning (DL) model based on improved YOLOv8 for the detection of the position and quantity of grouting sleeves in prefabricated concrete (PC) walls of assembled buildings. The model achieves efficient verification of grouting sleeves by combining target detection with region counting, effectively solving the problems of low efficiency, high cost, and lagging management of traditional manual inspection. In addition, the developed detection scheme can cooperate well with the production process and will not interfere with the production progress. In addition, the developed method has good scalability and can be further extended to the detection of other embedded parts in PC in the future. This study can be applied to the production line of PC factories to improve the quality inspection capabilities of enterprises, improve inspection efficiency and accuracy, reduce labor costs, and realize real-time and intelligent quality management, which has broad application prospects.