An improved YOLOv8-based method for concrete surface crack detection

The emergence and development of cracks in concrete structures can seriously threaten the safety of construction structures and increase the risk of deformation or even collapse of the building structure. However, existing inspection methods (manual inspection, machine vision technology and infrared imaging) have corresponding technical shortcomings, such as low inspection accuracy, high labour and equipment costs and inability to adapt to harsh inspection environments. Recognising concrete cracks using the YOLO algorithm in deep learning technology is one of the commonly used and effective means to avoid the above drawbacks. However, existing recognition methods based on the YOLO algorithm are unable to accurately recognise traces of water, vegetation, vineman, moss, mud and paint on the concrete surface, which reduces the accuracy of detection. To solve these problems and further improve the accuracy and rate of detection, an improved YOLOv8 algorithm is proposed in this study. We propose the BRSKA module for feature extraction and combine it with the EMA module. The improvement of the improved YOLOv8 algorithm mainly involves the convolutional layer, BRSKA module, EMA module and feature fusion and regression layer in the network structure. The experimental results showed that the improved YOLOv8 algorithm can improve the recognition accuracy compared with the current mainstream algorithms.