Research And Application Of A Lightweight Neural Network For Household Garbage Classification

At the end of 2018,215 million tons of garbage have been processed at home.About 35% of the unclassified garbage is recyclable resources.At the same time,the wet garbage in the unclassified garbage will cause incomplete combustion and produce a large amount of harmful gas.Considering the above factors,garbage classification has become more and more important.With the introduction of garbage classification policy in major cities,how to classify garbage correctly is a problem that every needs to face.In order to conveniently guide garbage classification,a lightweight neural network model that can run on the Android mobile terminal is trained.First of all,compare the existing garbage classification data sets,and select the Huawei data set as the basic data set.Using expansion methods such as crawlers,the number of samples has been increased from 14,683 to 20,560,and the number of garbage classifications has been increased from 40 to 42,,in order to solve the problem of insufficient data.Secondly,on the basis of transfer learning,a variety of classic network models are used to train on the garbage data set.At the same time,we compare the difference between different optimizers,loss functions,transfer learning and non-transfer learning.Then select the better network model as the teacher network to perform distillation operation on Mobile Net.It was found that the network result using the teaching assistant knowledge distillation was the best,and the test accuracy reached 82.4%.Finally,use the Py Torch Mobile framework to apply the model to Android phones.The system can call the camera to take a picture or select a picture from an album for recognition operation.Based on the actual application of garbage classification,this research trained a small network model based on transfer learning and knowledge distillation,and successfully applied it to the Android mobile,which can provide help for daily garbage classification work.