基于新型空间注意力机制和迁移学习的垃圾图像分类算法

随着我国各级政府大力推动垃圾强制分类,分类回收各环节中实现标准化、自动化的垃圾分类识别需要适合云端部署的高准确率、低延时要求的细粒度图像分类模型.本文发挥深度迁移学习的优势建立了一套端到端的迁移学习网络架构GANet (garbage neural network);针对垃圾分类中类别易混淆、背景干扰等挑战,提出一种新型的像素级空间注意力机制PSATT (pixel-level spatial attention).为克服类别多和样本不平衡挑战,提出使用标签平滑正则化损失函数;为改善收敛速度以及模型稳定性与泛化性,提出了阶梯形OneCycle学习率控制方法,并给出了结合Rectified Adam (RAdam)优化方法和权重平滑处理技术的组合使用策略.实验使用了"华为云人工智能大赛.垃圾分类挑战杯"提供的按照深圳市垃圾分类标准标注的训练数据,验证了GANet在垃圾分类问题中的显著效果,获得了全国二等奖(第2名);同时,提出的PSATT机制优于对比方法,且在不同主干网络架构上均得到了提升,具有良好的通用性.本文提出的GANet架构、PSATT机制和训练策略不仅具有重要的工程参考价值,也具有较好的学术价值.

Classification algorithm of garbage images based on novel spatial attention mechanism and transfer learning

As governments at all levels in China have started to promote mandatory garbage classification,in order to meet the standardization and automated garbage classification in all aspects of classification and recycling need a fine-grained image classification model suitable for cloud deployment with high accuracy and low latency.This article takes advantage of deep transfer learning to establish an end-to-end transfer learning network architecture GANet(garbage neural network).Aiming at the challenges of category confusion and background interference in garbage classification,this paper proposes a new pixel-level spatial attention mechanism PS ATT(pixel-level spatial attention).In order to overcome the challenges of multi-class and sample imbalance,this paper proposes a label smoothing regularization loss function.In order to improve convergence speed,model stability and generalization,this paper proposes a stepped OneCycle learning rate control method,and gives a combined use strategy combining Rectified Adam(RAdam) optimization method and stochastic weight averaging.Experiments used the training data which are marked by the Shenzhen garbage classification standard and provided by the "Huawei cloud artificial intelligence competition·garbage classification challenge cup",and verified the significant effect of GANet in the garbage classification problem,and won the national second prize(2 nd place).At the same time,the proposed PSATT mechanism is superior to the comparison methods with improvement on different backbone network architectures,and has good versatility.The GANet architecture,PSATT mechanism and training strategies proposed in this paper not only have important engineering reference value,but also have good academic value.