A Self-Supervised Power Equipment Defect Detection Algorithm Based on Multi-Level Feature Fusion

To fully leverage the massive amounts of unlabeled video surveillance image data from the power grid and address the issues of poor feature representation in single-layer features and low model recognition accuracy under limited labeled power equipment defect samples, we propose a self-supervised power equipment defect detection algorithm based on multi-level feature fusion. This algorithm first designs a feature fusion network that effectively integrates local and global multi-level features from shallow, intermediate, and deep layers of images, applicable to both CNN and Transformer architectures. Secondly, it introduces a masked completion self-supervised learning task to perform pretraining on the massive unlabeled power grid image data. Finally, based on the pre-trained power grid visual foundation model, finetuning is carried out for downstream object detection tasks to complete the training of the power equipment defect detection model. Experimental results show that compared to related single supervised object detection algorithms, this algorithm improves accuracy by $6.84 \%$ on the power equipment defect detection dataset, demonstrating better recognition accuracy and generalization.