Research on Power Target Detection Model Compression Technology Based on Knowledge Distillation

With the continuous progress of power system technology and the improvement of information level, the demand for power image data processing and analysis is increasing day by day. In this paper, knowledge distillation technology is used to transfer the rich knowledge of large complex models to small models. By allowing small models to learn and mimic the predictive behavior of large models, we achieve model compression while significantly improving the recognition accuracy of small models in power target detection tasks. In addition, model compression also brings a significant increase in inference speed, which is of great significance for real-time power system monitoring and fault diagnosis.