Design And Application Of Garbage Classification Algorithm Based On Neural Network Compression

The composition of domestic waste is complex and diverse,which has a great impact on human health and ecological development.In view of the current situation,various cities in China have begun to vigorously implement waste classification measures.Due to the various types of garbage,the accurate classification of garbage has become a major problem.At present,in the garbage classification,it is still in the state of manual sorting through the naked eye.It is not only inefficient,but also harmful to people’s health.With the continuous progress and development of deep learning,the research on garbage classification based on deep learning has important practical significance.In this paper,a pruning algorithm based on improved attention mechanism is proposed,and a device for automatic garbage classification is designed.At the same time,an automatic garbage classification system is designed using pyqt5.The feasibility of the system integration is verified by experiments.The main work of this paper includes:(1)A channel pruning algorithm based on attention mechanism is proposed.The improved hybrid attention mechanism module is used to generate weights for different channels in the same layer,which is used as the channel importance evaluation standard to guide the network to prune channels.Experiments show that the hybrid attention module proposed in this paper can significantly improve the network performance,and the pruning algorithm based on this attention mechanism can effectively compress the network while ensuring less accuracy loss.(2)In order to ensure the accuracy of the garbage recognition model,this paper filters,expands and enhances the existing public data sets Trash Net and garbage classify,and constructs the domestic garbage data set.(3)Based on NVIDIA Jetson Xavier NX,an automatic garbage can classification device is designed and manufactured,and a visual garbage classification operation interface is designed by pyqt5.The pruned model proposed in this paper is transplanted to the device for experimental test,and the overall measured classification accuracy can reach 91.5%.