融合注意力机制的垃圾分类迁移学习方法  

Garbage Classification Transfer Learning Method Based on Attention Mechanism

作　　者：柯健[1] 王敏[1] 张量[1] KE Jian;WANG Min;ZHANG Liang(School of Computer Engineering,Suzhou Vocational University,Suzhou Jiangsu 215104)

机构地区：[1]苏州市职业大学计算机工程学院,江苏苏州215104

出　　处：《软件》2021年第11期172-177,共6页Software

摘　　要：针对目前城市生活垃圾依赖人工分类效率低下,造成环境污染以及资源浪费等问题,提出了一种融合通道空间注意力机制的垃圾分类迁移学习方法。通过迁移学习,减少了垃圾分类模型的训练时间,通过改进Res Next101 32x16d wsl模型,增加注意力机制,提高了模型对垃圾特征的提取能力,使用Focal Loss代替标准交叉熵损失函数,解决了数据集样本不平衡的问题,提升了模型分类精度。通过实验证明,在华为垃圾公开数据集上,该模型对40小类的垃圾进行分类,其精度可以达到92.6%,表明模型有较强的泛化能力。Aiming at the problems of low efficiency,environmental pollution and waste of resources caused by the manual classification of municipal solid waste,a transfer learning method of garbage classification based on channel-spatial attention mechanism is proposed.Through transfer learning,the training time of garbage classification model is reduced,by improving the ResNext 101 32 x16 d wsl model and adding attention mechanism,the ability of garbage feature extraction is improved.Using Focal Loss instead of standard cross entropy loss function, the problem of sample imbalance in data set is solved,and the classification accuracy of the model is improved.Experiments show that the model classifies 40 categories of garbage on the Huawei garbage public data set,and its accuracy can reach 92.6%,which shows that the model has strong generalization ability.

关 键 词：垃圾分类 迁移学习 注意力机制 损失函数 残差网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]