基于多模型决策融合的可回收垃圾分类方法  

Classification Method of Recyclable Waste Based on Multi-modal Decision Fusion

作　　者：张露 理华[1] 崔杰[1] 王晓东[1] 肖灵[1] ZHANG Lu;LI Hua;CUI Jie;WANG Xiao Dong;XIAO Ling(Ultrasonics Lab Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190,China;University of Chinese Academy of Sciences,School of Electronic,Electrical and Communication Engineering,Beijing,100049,China)

机构地区：[1]中国科学院声学研究所超声学实验室,北京100190 [2]中国科学院大学电子电气与通信工程学院,北京100049

出　　处：《网络新媒体技术》2022年第6期48-56,共9页Network New Media Technology

摘　　要：为解决现存可回收垃圾分类方法较为低效,垃圾不能及时处理的问题,本文提出了一种基于多模型决策融合网络并实现了准确的垃圾分类。构建的决策融合网络以垃圾图像作为输入,选取经典神经网络Googlenet、VGG19_BN、Resnet18分别作为3个决策模型,融合3个决策模型的决策分类结果作为最终分类结果,实现更为可靠、精确的可回收垃圾分类。模型的训练还加入了迁移学习与学习率衰减的技巧。经过数据集验证,与其他可回收垃圾分类深度学习方法相比所提出的方法实现了更高的可回收垃圾分类准确率,其在数据集上的测试准确率达到97.67%,同时与单模型决策网络的比较结果验证了本文所提多模型决策融合方法的有效性。This research proposes an intelligent waste classification approach based on a multi-model decision fusion network to address the issue that the existing waste classification method is inefficient and cannot deal with rubbish in a timely manner.The proposed decision fusion model starts with a trash image as input,then chooses classic neural networks Googlenet,VGG19_BN,and Resnet18 as decision models,and then fuses the decision classification results of the three decision models to provide the final classification results.Decision fusion is a concept that enables for more accurate and dependable recycling trash sorting.Transfer learning and learning rate decay are also included in the model’s training.Compared with other deep learning methods for recyclable waste classification,the proposed method achieves higher classification accuracy,and its test accuracy on the data set is 97.67%,In addition,the comparison with the single-model decision-making network verifies the validity of the multi-model decision-making fusion method.

关 键 词：可回收垃圾分类 决策模型 迁移学习 学习率衰减 决策融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] X799.3[自动化与计算机技术—计算机科学与技术]