A Road Pothole Detection Algorithm for Small Targets Based on Improved YOLOv8

To enhance the performance of the YOLOv8 model in road pothole detection for small targets, this paper proposes two improvement methods: a base model improvement method based on attention mechanisms and an advanced model improvement method involving an enhanced loss function. First, three types of attention mechanisms—SE, CBAM, and ECA—were introduced after the fourth C2f module in the Head network of YOLOv8 to improve the model’s ability to extract key information. Second, to further optimize the bounding box regression accuracy in small target detection tasks, this paper proposes a loss function, NWD-GIoU, which combines the Normalized Wasserstein Distance (NWD) and Generalized IoU (GIoU). The balance between the two is controlled by the parameter α, where α represents the weight coefficient of the GIoU component in the loss function, and (1-α) is the weight coefficient for the NWD component. Experimental results with attention mechanisms indicate that the SE module performed the best among the three, improving the model’s performance by 0.4% and 0.2% in mAP50 and mAP50:95, respectively, compared to the original YOLOv8 model. The loss function experiments show that when $\alpha=0.4$, the model achieved the best precision, with an improvement of 2.8% and 0.8% in mAP50 and mAP50:95, respectively, compared to the original YOLOv8 model. Thus, both the introduction of the SE attention mechanism and the use of the improved NWD-GIoU loss function enhanced the model’s performance in road pothole detection for small targets. However, when combining the SE attention mechanism, which performed best in the attention mechanism experiments, with the improved NWD-GIoU loss function, the model’s performance increased by 1.3% and 0.1% in mAP50 and mAP50:95, respectively, compared to the original YOLOv8 model. This improvement was not as significant as using the NWD-GIoU loss function alone, suggesting that the proposed NWD-GIoU loss function optimizes the model more effectively. Further exploration is needed to balance and coordinate the relationship between attention mechanisms and improved loss functions in the model’s optimization strategy.