基于机器视觉的路边垃圾分类系统

Roadside waste classification system based on machine vision

基于机器视觉技术设计制作了路边垃圾分类系统,主要包括硬件电路和垃圾识别模型。改进了现有多注意力模块,使模型更轻量化,采用H-Swish激活函数提高识别准确率,然后在公开的华为垃圾分类比赛数据集上进行仿真实验。结果表明:该方法识别准确率达到87.35%。针对自建数据集数据量少、过拟合严重的问题,采用迁移学习的方法,将在华为数据集上训练完成的模型参数进行迁移,在自建数据集上继续训练。最后,将模型部署到树莓派4B上,在制作的实物平台上进行测试。结果表明:该系统平均一次回收需要2 s,可以有效地进行路边垃圾识别分类。

A roadside waste classification system is designed and manufactured based on machine vision technology,mainly includes hardware circuit and waste identification model.The existing multi-attention module is improved to make the model more lighter,and the H-Swish activation function is used to improve the recognition accuracy,then the simulation experiment is carried out on the public Huawei waste classification competition dataset.The results show that the recognition accuracy of this method reaches 87.35%.Aiming at the problem of small amount of data and serious over fitting in the self-built dataset,a migration learning method is adopted to migrate the model parameters trained on the Huawei dataset and continue to train on the self-built dataset.Finally,the model is deployed on Raspberry Pi 4B and tested on the physical platform.The results show that the average onetime recycling of the system takes 2 s,which can effectively identify and classify roadside waste.