Research On Garbage Classification And Recognition Algorithm Based On Lightweight CNN

The acceleration of urbanization,Make the amount of garbage rise rapidly.Correctly sorting garbage and improving recycling efficiency can not only protect the environment but also save resources,Since 2019,Many cities have gradually introduced mandatory garbage classification guidelines.However,At present,there are not many intelligent garbage sorting applications,The domestic intelligent garbage classification research starts late.The current garbage classification algorithm has the disadvantages of too much network parameters,not easy to transplant into mobile devices,and poor adaptability in actual scenarios.This paper studies six different network models,And experimented on a common dataset Trashnet,The lightweight network MobileNetV2 and EfficientNetB0 with both accuracy and parameter are selected as the baseline network.Besides using Trashnet data sets,A garbage image data set was also established(Trash2020),It contains 40 kinds of garbage.For MobileNetV2 network classification accuracy,To reduce the network size and prevent overfitting,Using migration learning strategies to train MobileNetV2 networks ImageNet large-scale data sets,Build the trained network fusion attention mechanism into the A-MobileNet network,Then migrate to the Trashnet data set and Trash2020 data set for fine-tuning.The experiment shows that the accuracy of the A-MobileNet model is 8%higher than that of the MobileNetV2 model,The classification accuracy reached 90.In practice,the accuracy of the model is high,In order to improve the accuracy of garbage classification,A garbage image classification method based on local receptive field expansion D-EfficientNet model is proposed,Because cavity convolution approval expands the local receptive field without adding parameters,Hence adding a cavity convolution to the EfficientNetB0 network to expand the final feature extraction,Improving the accuracy of the model without increasing the number of parameters,and Disout algorithm is used in training to prevent overfitting.Experimental results show that the classification accuracy of D-EfficientNet network is 92.2%,3%more than EfficientNetB0.The experiment showed that the D-EfficientNet network improved by 2.2%and 1.56%compared with the A-Mobilenet Trash2020 dataset.Compared with other literature algorithms,the classification accuracy of the D-EfficientNet model also improved.The results show that the D-EfficientNet model has better generalization and robustness in garbage image classification.