AUTHOR=Xing Yan , Han Xu , Pan Xiaodong , An Dong , Liu Weidong , Bai Yuanshen 

TITLE=EMG-YOLO: road crack detection algorithm for edge computing devices

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 18 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2024.1423738

DOI=10.3389/fnbot.2024.1423738

ISSN=1662-5218

ABSTRACT=Road cracks seriously shorten the service cycle of the road. Inefficiency and high cost of manual detection methods. Currently the YOLOv5 model has made some progress in road crack detection. However, there are issues when deployed on edge computing devices . The main problem is that edge computing devices are directly connected to sensors . Device collects noisy, poor quality data. This problem puts additional computational burden on the model, potentially impacting its accuracy. To address these problems, this article proposes a novel road crack detection algorithm . The algorithm is named EMG-YOLO . Firstly, in YOLOv5, Efficient decoupled header is introduced to optimise the Head. The purpose of this approach is to separate the classification task from the localization task. Each task can focus on learning its most relevant features. It reduces the model's computational resources and time by a large amount . Models can achieve faster convergence rates. Secondly, IOU loss function in the model is upgraded to MPDIOU loss function. It works by minimising the top left and bottom right point distances between the predicted bounding box and the actual labelled bounding box . The MPDIOU loss function can solve the problem of the current YOLOv5 model's relatively complex computation and high computational burden. Finally, GCC3 module replaces the traditional convolution . It performs global context modelling with the input feature map to obtain global context information. Model detection capabilities are enhanced on edge computing devices. The experimental results show that the improved model has improved in all parameter indicators compared with the current mainstream algorithm. EMG-YOLO improves the accuracy of YOLOv5 model by 2.7%, mAP(0.5) and mAP(0.9%) by 2.9% and 0.9%, respectively. The new algorithm also outperforms the YOLOv5 model in complex environments in edge computing devices.