Research on Lightweight Convolutional Neural Network in Garbage Classification

Abstract With the rapid development of society, more and more garbage consumables are produced. How to better recycle and use “garbage” has become a widespread concern. In this paper, a lightweight garbage classification model (GroupAtten_ MobileNet, GA_MobileNet) is designed based on the convolutional neural network ResNet-50. Reduce the amount of Params, FLOPs and memory consumption by using deep convolution and group convolution; use the channel attention mechanism to increase the accuracy of the model; Use the model fine-tuning method to further reduce the model’s Params, FLOPs, and memory consumption. This paper divides the recyclable garbage into 6 categories. Through an experimental comparison on the garbage data set, it is found that the amount of GA_MobileNet Params proposed in this paper is 280 times less than ResNet-50; the FLOPs amount is reduced by 31.7 times; the accuracy rate Increased by about 3.1%; operating time reduced by 50%;At the same time, compared with the same type of lightweight models Mobile Netv 2, Squeeze Net and shuffle Net V1 also has a clear advantage. Finally, it is concluded that GA_MobileNet is a garbage classification model suitable for lightweight convolutional neural networks, which is suitable for recycling of garbage, protecting the environment and benefiting the country and the people.