Knowledge Distillation Based Lightweight Satellite Video Motion Target Detection Algorithm

Existing satellite video target detection models have problems such as large number of parameters, slow detection speed, and difficult to be deployed at the edge of mobile devices. The lightweight detection model LDSFNet, obtained by model compression method, can greatly reduce the computational complexity and storage space occupation. In order to solve the problem of degradation of detection performance of LDSFNet, we design a knowledge distillation framework, which is used in training to optimize the model obtained. Compared with the original complex model, the number of parameters and computation amount are significantly reduced, the accuracy and the recall show only a small decline after adopting the knowledge distillation method proposed in this paper. The lightweight model LDSFNet finally realizes a balance between the model in terms of speed and performance.