| Abstract |
In the context of smart city scenarios, identifying manhole cover hazards faces challenges such as large inter-class differences and small intra-class differences among hazard categories, leading to low detection accuracy, false detections, and missed detections. This paper proposes a fine-grained manhole cover hidden danger detection method based on YOLOv8 and transfer learning. Firstly, to address the issue of large inter-class differences and small intra-class differences among hidden danger categories, the ACmix self-attention mechanism module was added to the base model. Then, to solve the problem of uneven distribution between difficult and simple samples, the Focaler-IoU loss function was introduced in the regression loss part. Finally, to address the issue of the small amount of data in the existing dataset, data augmentation was performed through methods such as scale variation and noise addition, and the augmented dataset was labeled using the LabelImg software to construct a new dataset for model training. Experimental results show that the improved YOLOv8 achieves an average precision (mAP) of 99.1% in manhole cover hidden danger detection, which is an improvement of 0.4% compared to the original YOLOv8 algorithm. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025. |
