Research And Application Of Garbage Classification Algorithm Based On Deep Transfer Learnin

The explosion of garbage brought about by the improvement of living standard of our residents not only causes serious pollution to the environment,but also endangers the health of residents.Although China has gradually started to implement the policy of waste separation,there are still problems that residents’ consciousness of waste separation is late and factories still adopt manual assembly line to separate waste.Nowadays,with the unprecedented progress of deep learning technology in the field of computer vision,migration learning technology,which is closely related to it,has also gained rapid development.It provides a new solution to efficiently deal with the problem of waste classification.Therefore,this paper designs and applies a garbage classification algorithm model based on the deep learning approach combined with the migration learning technology,balancing the two indicators of recognition accuracy and hardware consumption.The main research contents of this paper are as follows.First,based on the study of the current national standards for garbage classification in China,a garbage data set is established.The garbage pictures are collected by data crawling,and the collected pictures are screened,sorted and labeled.A garbage dataset with 4 major categories including 210 subcategories is established,and the image samples are also pre-processed and data enhanced to improve the data guarantee for deep migration learning in this paper.Second,after establishing the dataset,different convolutional neural networks are constructed based on deep learning to compare the experiments,and the model with the best performance is used as the backbone network.Based on the backbone network,we build the(Attention Module-Adaptive Pooling,AM-AP)network which incorporates the attention mechanism and the adaptive pooling module.The experimental results show that the recognition accuracy of the model can reach87.82%,which is 4.19% better than the benchmark model,while the hardware load only increases by about 1%,and the model will be used as the base model for subsequent deep migration learning.Third,on the basis of AM-AP network,deep learning and migration learning techniques are combined to build(Garbage Deep Transfer Network,GTDN)network,including the proposed progressive chunk migration strategy and stepwise fine-tuning network.Through experimental analysis,the best chunked migration strategy with fine-tuning network is compared,and the final recognition accuracy reaches 97.86%,while the training time decreases by 53.63% and the video memory load is reduced by18.17%.Finally,using GTDN model,a visual application is designed and written to realize the functions of garbage fast reading,identification and guide,reflecting the high practicality of the algorithm in this paper.