基于深度迁移学习的垃圾分类研究

Deep Transfer Learning-Based Waste Classification Research

针对垃圾分类人工检测环境差,易出错,难度大,效率低的问题,提出一种利用深度迁移学习对生活垃圾分类的方法。首先,构建垃圾分类的图像数据集,同时数据增强,其次,搭建深度卷积神经网络ResNeXt和MobileNetV2,微调网络迁移参数以适应垃圾分类任务,最后,在基于深度迁移学习的卷积神经网络下,探索了网络冻结层数和学习率对不同量级的网络结构造成的影响。结果表明,ResNeXt受到学习率的影响更强,MobileNetV2受到网络冻结层数的影响更多,两者的最佳网络冻结层分别是50层和80层,最佳学习率分别是0.0003和0.0001,有效提升模型准确率,实现了对多种常见垃圾的有效分类。

To address the problems of poor manual detection environment, error-prone, difficulty and low efficiency of garbage classification, a method of domestic garbage classification using deep transfer learning is proposed. Firstly, image datasets for garbage classification are constructed while data augmentation, secondly, deep convolutional neural networks ResNeXt and MobileNetV2 are built to fine-tune the network transfer parameters to suit the garbage classification task, and finally, the effects of network freezing layers and learning rate on the network structure caused by different magnitudes are explored under the convolutional neural networks based on deep migration learning. The results show that ResNeXt is more strongly influenced by the learning rate and MobileNetV2 is more influenced by the number of network freeze layers, and the best network freeze layers for both are 50 and 80 layers, respectively, and the best learning rates are 0.0003 and 0.0001, respectively, which effectively improve the model accuracy and achieve the effective classification of many kinds of common garbage.