深度迁移学习与注意力机制的垃圾图像分类方法

Garbage Image Classification Based on Deep Transfer Learning and Attention Module

垃圾分类对节约资源和改善环境有着重大的意义。针对日益增长的消费能力带来的垃圾种类的增长,提出基于深度学习神经网络和迁移学习并引入注意力机制的垃圾分类图像识别方法。首先,建立以可回收垃圾、厨余垃圾、有害垃圾和其他垃圾四大分类为基础,其中包括210种子分类的二级分类数据集;其次,采用自训练加迁移学习融合的方式,先搭建自训练卷积神经网络,后建立VGG16、ResNet50和ResNeSt50卷积神经网络,迁移同构模型下的预训练特征模型,把两个网络提取的特征融合,再添加基于CBAM注意力机制的改进模型,最后再接入微调网络再训练。分析得出最好的垃圾分类模型。实验数据表明,论文方法对比非迁移学习网络,时间消耗平均节约了29.4%,模型准确度平均提升8.06%,准确度最高达到92.2%。该方式可以显著地提升垃圾分类自动化的效率。

Garbage sorting is of great significance to save resources and improve the environment.In response to the growth of garbage types brought by the increasing consumption power,a garbage classification image recognition method based on deep learn-ing neural network and migration learning and introducing attention mechanism is proposed.First,the four major classifications of recyclable waste,food waste,hazardous waste and other waste with national standards are collected and established,including the secondary classification data set with 210 seed classifications.Second,a self-training plus migration learning fusion approach is used to first build a self-training convolutional neural network,followed by VGG16,ResNet50 and ResNeSt50 convolutional neural networks,migrating homogeneous models under the pre-trained feature models are fused with the features extracted from the two networks,and then an improved model based on the CBAM attention mechanism is added,and finally the fine-tuned network is ac-cessed for retraining.The analysis yields the best garbage classification model.The experimental data shows that the method in this paper saves an average of 29.4%in time consumption and improves the model accuracy by 8.06%on average compared to the non-migratory learning network,with a maximum accuracy of 92.2%.The approach can significantly improve the efficiency of gar-bage classification automation.