A Electric Power Scene Object Detection Algorithm Based on Meta-learning and Category Decoupling

With the large-scale application and practical improvement of intelligent inspection in electric power scene, the ability to identify equipment defects has become increasingly important. Currently, the mainstream algorithms for equipment defect identification primarily rely on established techniques such as YOLO, CenterNet, etc. These algorithms perform exceptionally well on public datasets. However, in terms of the imbalance of class and the conflict between conceptual categories for equipment defect identification tasks, there is still room for optimization in these algorithms. In this work, we propose a YOLO-paradigm algorithm combined with meta-learning, which resolves issues of class imbalance and conceptual conflict between categories by decoupling the categories of detection head and providing categorical instance support through meta-learning. This algorithm is implemented through end-to-end training, and during the inference stage, the weighted coefficients obtained from the meta-learning module are fixed, reducing computational overhead. On the equipment defect dataset we have made, the proposed method in this work demonstrates better performance than the baseline model, especially for categories with few samples.