An application case of object detection model based on Yolov3-SPP model pruning

In order to reduce environmental pollution and accelerate the sustainable utilization of resources, it is very important to classify municipal solid waste. The traditional manual sorting method not only has high labor costs, but also can cause physical harm and threaten the health of practitioners when it comes into contact with hazardous waste. Therefore, it is of great research significance to improve the autonomy, intelligence and practical application of waste sorting. In the specific application of garbage classification, it is restricted by the hardware of edge devices, and it is necessary to meet the requirements of detection accuracy and reduction of computing power consumption at the same time. Therefore, this paper adopts the light YOLOv3-SPP network to classify garbage. Specifically, this paper first builds a garbage detection dataset, and then uses the dataset to train the YOLOv3-SPP network to form a YOLOv3-SPP garbage detection model. Next, model compression is performed on the YOLOv3-SPP garbage detection model using channel pruning and layer pruning, and the compressed model can be directly deployed on the low-computing edge computing device carried in the robot that automatically sorts and recycles garbage, so as to realize the miniaturization and large-scale practical application of the garbage sorting robot. The experimental results show that after model compression, the size and inference time of the model drop by 90.77% and 72.93%, respectively, while the mAP of the model only drops by 3.92%. It shows that the constructed garbage detection model meets the application requirements of real-time detection on edge computing devices with low computing power, and has practical application potential.