Switch object detection based on knowledge distillation for real time edge computing

With the development of Information Technology, real-time equipment target detection on the edge side has become increasingly important, particularly for the automatic control of electrical devices. Although existing lightweight object detection models have achieved a certain balance of accuracy and speed to a certain extent, they are generally not effective in dense small target recognition, and strengthening the detailed features directly through the method of feature fusion will destroy the lightweight attribute. Some existing object detection models which are suitable for multiple scenes, have good generalization performance and high precision, but cannot meet the reasoning requirements of end-to-side deployment in speed. Therefore, aiming at the actual situation that the data distribution of vertical scene is relatively simple, this paper successfully learns the lightweight model for target detection of distribution cabinet switches by means of knowledge distillation, using CenterNet model based on hourglass-104 as teacher and CenterNet model based on hourglass-52 as student. The experimental results demonstrate that the model can quickly and accurately detect the switching components of the distribution cabinet at the end side. Compared with the existing model, the detection speed is about twice as fast while maintaining high-precision detection.