Research On Recyclable Garbage Detection Algorithm Based On Deep Learning

Garbage classification is an important way to save renewable resources and reduce environmental pollution.Since the outbreak of the COVID-19 at the end of 2019,the spread of the epidemic has brought great obstacles to the garbage classification work in China.Deep learning technology has gradually become a hot technology under the rapid development of the third wave of artificial intelligence technology,and has been widely used in various fields.In order to improve the efficiency of recyclable garbage sorting and recycling,reduce manual participation in sorting or completely replace manual sorting,this paper applies the target detection algorithm in deep learning technology to the effective identification and positioning of recyclable garbage,so as to achieve rapid recycling of recyclable garbage.Efficient automatic classification.The main research work of this paper is as follows:（1）Make an exclusive data set for six types of recyclable garbage,and complete annotation of data set.（2）According to the average accuracy and detection speed of recyclable garbage target detection,the network models of SSD,YOLOv3 and YOLOv5 m are constructed and compared and verified by experiments.The experimental results show that the YOLOv5 m algorithm has higher detection accuracy for recyclable garbage data set.（3）Aiming at the problem that the YOLOv5 m algorithm has poor discrimination of occluded targets and relatively slow detection speed,an improved lightweight YOLOv5 m recyclable garbage target detection algorithm is proposed.Mobile Net V3-Large is used for feature extraction to reduce the amount of network parameters and model size;DIo U-NMS is used for prediction frame screening to improve the recognition of occluded targets;CIo U-Loss is used as a position loss function to speed up the convergence of the model.The results of experiments on the recyclable garbage data set using the improved algorithm show that the improved YOLOv5 m algorithm can significantly improve the recognition of occluded targets,while ensuring a high average detection accuracy,the amount of parameters is reduced by 76%,the model size is reduced by 74.1%,the computing power requirement is reduced by 80.2%,and the inference speed on the CPU is increased by72.8%,which can meet the needs of deployment in embedded and other devices.