Research on road crack detection in complex background based on MFF-YOLO

Road crack detection occupies a very important position in highway maintenance. With the development of deep learning, target detection has come into the public's view, facing the road crack data with irregular shape, complex pavement background and huge amount of data, seeking high precision, high efficiency and low cost target detector has become a popular research direction nowadays. In this paper, an improved MFF-YOLO model is proposed using YOLOv8 as the base model. By adding the BOTNet module to the backbone network, the pavement disease detection accuracy under complex background is improved; the idea of BIFPN is borrowed in the neck network, and the two-way cross-scale connection and weighted splicing operation are utilized to improve the reuse rate of the feature information and the detection effect on the small target disease; and the EIOU is used to replace the CIOU as the loss function of the detection frame regression in order to improve the model detection accuracy. The results show that the F1 value of the improved model proposed in this study reaches 69.4% and the mAP value reaches 73.7%, which are 1.3% and 3.3% higher than that of YOLOv8, respectively, and its detection accuracy is remarkable and able to cope with the detection of road cracks in complex backgrounds.