基于高效的动态网络垃圾图像分类模型研究  

Research on Efficient Dynamic Network for Garbage Image Classification

作　　者：黄日辰 陈晓龙[1] HUANG Richen;CHEN Xiaolong(Jinhua Polytechnic,Jinhua 321007,China)

机构地区：[1]金华职业技术学院,浙江金华321007

出　　处：《金华职业技术学院学报》2023年第3期60-67,共8页Journal of Jinhua Polytechnic

摘　　要：环境问题是一个非常严峻的全球性问题,而垃圾分类是一种保护环境的有效措施。针对现有的垃圾分类网络无法根据样本分类的难易程度动态平衡计算资源和识别准确率低等问题,提出了一种高效的动态垃圾分类网络DynamicGarbageNet。该网络基于MobileNetV2设计了一种动态的轻量化主干网络,并由三个深度依次递增的子网络组成。利用迁移学习技术优化模型参数,进一步提高精度。实验结果表明,DynamicGarbageNet在华为云垃圾分类数据集上的准确率为95.7%,速度为90FPS,均高于MobileNetV2。Environmental problems are a very serious global issue,and waste classification is an effective measure to protect the environment.In this paper,an efficient dynamic garbage classification network DynamicGarbageNet is proposed to address the problem that existing garbage classification networks cannot dynamically balance computational resources and accuracy according to the difficulty of samples.Based on MobileNetV2,a dynamic lightweight backbone network is designed,which consists of three sub-networks with increasing depth in sequence.The transfer learning technique is used to optimize the model parameters and further improve the accuracy.Experimental results show that DynamicGarbageNet achieves a garbage classification accuracy of 95.7% and a speed of 90 FPS on the Huawei Cloud garbage classification dataset,both surpassing the performance of MobileNetV2.

关 键 词：垃圾分类 迁移学习 动态网络 高效网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]