SD-YOLO: A lightweight manhole cover defect detection method with diffusion model data augmentation

Manhole cover maintenance is a crucial aspect of urban infrastructure management, as defective manhole covers pose significant risks to pedestrian safety and traffic operations. However, deep learning-based detection methods face two key challenges in practical applications: 1) Due to the high cost of manual annotation and the rarity of defects, it is extremely difficult to obtain fine-grained annotated data for manhole cover defects. 2) Data augmented through conventional techniques such as flipping and rotating lacks the semantic and content diversity of specific defects, and cannot address the issue of data scarcity. The main challenges at this stage are how to generate high-semantic, high-quality defective samples, and how to design a lightweight model that accurately captures fine-grained defect features without sacrificing inference efficiency. To address these challenges, we propose a lightweight model named SD-YOLO for manhole cover defect detection under limited data. This model integrates two key innovations: 1) Stable Diffusion Data Augmentation (SDDA) strategy, which uses a LoRA fine-tuned stable diffusion model to generate virtual defective images with real semantics and diverse scenes; 2) Local-Global Fusion Attention (LGFA) module, which adaptively integrates local texture details and global structural features of defects to enhance fine-grained detection capabilities. Experimental results show that SD-YOLO achieves an F1 score of 83.9% on a self-built manhole cover defect dataset (MCDD), effectively alleviating the performance bottleneck of minority class detection caused by data imbalance.