Target detection on lightweight device based on Compressed YOLOv5s6

In recent years, with the development of deep learning and target detection, the accuracy of detection network is higher and higher, and the increase of network parameters and the decrease of inference speed. However, in actual application scenarios, the detection network needs to be deployed on some mobile or lightweight devices. To solve this problem, this paper proposes a method to compress the model. Based on YOLOv5s6 model, the channels with small weight are removed through sparse training and channel pruning, and then fix the model accuracy by knowledge distillation. Finally, the lightweight model Compressed YOLOv5s6 is obtained. The experimental result shows that the Compressed YOLOv5s6 model reduces 95.1% of the parameters, 30% of the inference speed and 90.2% of the model size compared with the original model, which is more suitable for the application of practical scenes.