Research On Garbage Classification Based On Deep Learning

Nowadays,with the rapid development of society and the increasing improvement of people’s living standards,the output of domestic waste has also increased year by year,which has brought severe challenges to urban management,environmental protection and resource recycling.In this regard,China continues to strengthen the promotion of waste classification.While implementing waste classification policies on the residential side,it is also promoting industrial upgrading on the factory side.At present,most domestic waste treatment plants are mainly manual sorting,which seriously affects the health of workers at the same time.How to effectively improve the efficiency of waste classification and recycling has become an urgent problem to be solved in the current society.With the vigorous development of deep learning technology,more and more practical problems have been solved,which also makes researchers see the possibility of using AI for garbage classification,and how to design a garbage classification algorithm with excellent classification performance has become a key to solve the garbage classification problem.Based on the above situation,Thesis focuses on the research of domestic waste classification technology based on in-depth learning.The main research contents are as follows:(1)Firstly,thesis investigates the existing domestic waste public data set,and finds that the current mainstream data set-huaweiyun domestic waste data set has the problem of data imbalance.In the preliminary training process,a small number of categories are easy to be fitted,which will affect the classification performance of the model as a whole.At the same time,it makes an in-depth study on the discrimination of various types in the classification process,It is found that there are similarities between different types of garbage,which is easy to be confused in classification and discrimination.thesis carries out further research based on the above two issues.(2)In order to solve the problem of data set imbalance,based on the existing balanced generation countermeasure network Bagan,thesis improves its defect of image generation instability,and proposes an enhanced balanced generation countermeasure network ebgan,which is used to generate pseudo pictures to expand the original data set.Experiments show that compared with Bagan,the FID score of ebgan generated images is reduced by 59.94,and the SSIM score is increased by 0.105,which verifies that the network has good image generation performance.At the same time,in the classification experiment based on resnext101 model,the accuracy of the data enhancement method based on ebgan network is 1.46% higher than that of the unused data enhancement method,and 0.71% higher than that of the traditional data enhancement method.It is verified that this method can effectively deal with the imbalance of data sets and improve the overall classification performance of the model.(3)In order to solve the problem that similar items are easy to be confused in the classification process,thesis improves the loss function,proposes a risk item constrained loss function,introduces the attention mechanism,and designs a hybrid attention module composed of multi-scale channel attention mechanism and autocorrelation spatial attention mechanism to improve the overall classification accuracy of the algorithm.The experimental results show that,The accuracy of the two methods in resnext101 model is improved by 0.59% and 1.09% respectively,and the overall accuracy is improved by 1.38%,which verifies the superiority of the proposed method.