Design And Research Of Domestic Garbage Classification Algorithm And Mobile Terminal System Based On Deep Learning

As a member of the earth,while enjoying the blessings of nature,we should also learn to protect nature and promote the sustainable development of natural resources.The treatment of domestic garbage is closely related to environmental protection.As city residents,we should do a good job in the classification of domestic garbage and contribute to the protection of the environment.However,as the types of domestic garbage continue to increase,there are differences in the classification standards of domestic garbage in different cities,which has become the biggest problem in the classification of domestic garbage at present.In this context,this thesis establishes a deep learning model for domestic garbage classification,and continuously improves and optimizes the model.Finally,the model is deployed on mobile devices to help people solve the problem of domestic waste classification.First,the performance evaluation index of the model is constructed,which includes four performance indexes: Accuracy,Params,FLOPs and Latency.At the same time,a data set for the classification of domestic garbage is established.The standard for establishing the data set in this article refers to the Guangzhou garbage classification policy,and the domestic garbage is divided into 4 major categories and 30 sub-categories,including common garbage in daily life.Secondly,based on four performance indicators,the performance of different models in the task of garbage classification is compared,and the GhostNet network is selected as the basic network of this article.In order to further improve the performance of the GhostNet network,a training strategy suitable for the GhostNet network is selected by comparing different transfer learning,loss functions and optimizati on algorithms.On this basis,the structure of the GhostNet network is improved.By introducing CBAM,the accuracy of the model is improved on the premise that other performance is almost unchanged.Then,after the structural improvement of the GhostNet ne twork,the GhostNet＿CBAM＿V2 network is selected for further optimization.Through analysis,it is found that there are a large number of parameters in the last fully connected layer of the GhostNet＿CBAM＿V2 network,occupying 31.3% of the entire network.In response to this situation,a corresponding compression strategy is proposed,and different compression coefficients are changed to compress the network.Through comparative experiments,it is verified that there is lots of parameter redundancy in the last fully connected layer of the network.In the end,the GhostNet＿CBAM＿V2＿80 compression network was selected.Under the premise that other performances are almost unchanged,the amount of network parameters is only 71.1% of the original.Then,after the model was determined,two different ways of App were designed to deploy the network on mobile devices.The App designed in this thesis is compatible with the Android operating system,which is a photo mode and a video mode to meet the usage habits of different users.Test the App on different Android devices and versions to verify that it has good compatibility.Finally,the research content is summarized and prospected.