IDD-YOLOv7: A lightweight and efficient feature extraction method for insulator defect detection

Aiming at the challenges of diversity, large scale variation and complex background in insulator defect images captured by UAVs, as well as the insufficient ability of existing detection algorithms in terms of leakage and false detection of small-size defects and adaptation to complex backgrounds, this study proposes a lightweight and efficient feature extraction method for insulator defects detection based on YOLOv7, named IDD-YOLOv7. First, we propose a Multi-Scale Channel Information Extraction Module (MCIE), enabling the model to effectively learn and utilize feature map information at different scales to address the challenge of significant scale variations in defect areas. Second, we propose a Contextual Global-Local Attention Module (CG-LA), allowing the model to consider both global context information and local details to tackle background interference issues. Additionally, we design a Focused Pure Convolution Feature Extraction Module (FPCFE) to enhance the model's focus on tiny insulator defects, effectively addressing the issues of missed detections and false positives in existing algorithms for small-sized defects. To improve the model's robustness and generalization ability, we apply data augmentation to the defect dataset. Finally, we utilize channel pruning and knowledge distillation strategies to compress the improved model, making it lightweight enough for deployment on UAV platforms with limited computational resources. The experimental results show that compared with the baseline model, IDD-YOLOv7 improves the AP50 in breakage-defect, drop-defect and flashover-defect by 10.3 %, 8.8 % and 10.8 %, respectively, and the mAP improves by 8.9 %, and the physical storage space occupies only 5.6MB, which is compared to YOLOv7 reduced by 69 %. Compared with the existing detection algorithms, IDD-YOLOv7 is not only able to detect various insulator defects quickly and accurately, but also has low leakage and false detection rates. Taken together, the proposed algorithm has significant robustness in insulator defect detection.