Road disease detection algorithm based on improved YOLOv8

In recent years, with the rapid development of highway transportation and the increasing prevalence of vehicles, road surface damage has become increasingly severe. This not only significantly increases road maintenance tasks but also raises the likelihood of traffic accidents, thereby affecting traffic efficiency and socio-economic development. Consequently, the demand for road damage detection is growing, and the requirements are becoming more stringent. Accurate and efficient road damage detection has become a pressing need. In real-world environments, road conditions are complex and variable, and are significantly affected by environmental factors such as lighting and shadows. These challenges make traditional image processing techniques less effective in detecting road damage. Since road damage often appears as fine cracks that occupy few pixels and have indistinct features, there is a need to develop more advanced detection models. To meet the high demands for accuracy and real-time performance in road damage detection, this paper innovatively proposes a deep learning model-YOLOv8 integrated with a Transformer-specifically designed for road damage detection. This model combines the powerful feature extraction capabilities of the Transformer with the fast object detection advantages of YOLOv8, aiming to more effectively identify road damage and achieve high-precision, high-efficiency detection. The model is suitable for road monitoring and maintenance in urban traffic, highways, and other scenarios, helping to improve road safety and transportation efficiency. Through the application of deep learning technology, we aim to bring revolutionary improvements to road maintenance and management. The model uses YOLOv8n as the base structure, and incorporates the Swin Transformer into the backbone feature network. The Swin Transformer employs a shifted window approach to reduce computational complexity and utilizes window interactions to achieve global transformation. Experiments were conducted on the RDD2022 dataset, and the results demonstrate that the proposed method effectively improves detection accuracy while maintaining good real-time performance. Additionally, the lightweight model is more easily deployable on mobile devices.