多尺度特征融合的轻量型垃圾分类方法  

Lightweight Garbage Classification Method Based on Multi-scale Feature Fusion

作　　者：高静 段中兴[1] 何宇超 GAO Jing;DUAN Zhong-xing;HE Yu-chao(College of Information and Control Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,西安710055

出　　处：《小型微型计算机系统》2023年第2期376-382,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(51678470)资助.

摘　　要：针对垃圾图像背景复杂,类别易混淆,现有垃圾分类模型难以兼顾高精度、小体积与低延时要求的问题,建立了一个多尺度特征融合的轻量型垃圾分类网络ML-Xception(Multiscale Lightweight Xception),提出多尺度特征提取模块,进行特征融合,提升特征信息的丰富性;在输出层使用锯齿状扩张卷积,增强了深层特征的提取能力;增加Dropout模块缓解过拟合问题,并对网络进行裁剪优化.在优化策略中,提出了预热与余弦退火相结合的学习率控制方法;使用Gridmask数据增强提高了数据的多样性.在“华为云人工智能大赛·垃圾分类挑战杯”提供的数据集上,分类准确率为97.3%,推理速率为25ms/张,在分类精度与推理时间等方面均优于其他模型,具有重要的工程应用参考价值.Aiming at the problem that the background of garbage image is complex and the classification is easy to be confused,and the existing garbage classification models are difficult to meet the requirements of high precision,small volume and low delay,a lightweight garbage classification network ML-Xception(Multiscale Lightweight Xception)based on multi-scale feature fusion is established,and a multi-scale feature extraction module is proposed to perform feature fusion and improve the richness of feature information.The sawtooth expansion convolution is used in the output layer to enhance the extraction ability of deep features.The Dropout module is added to alleviate the overfitting problem and optimize the network.In the optimization strategy,a learning rate control method combining preheating and cosine annealing is proposed.The use of Gridmask data enhancement improves the diversity of data.On the data set provided by′Huawei Cloud Artificial Intelligence Competition Garbage Classification Challenge Cup′,the classification accuracy is 97.3%and the reasoning rate is 25 ms/sheet,which is better than other models in classification accuracy and reasoning time,and has important engineering application reference value.

关 键 词：垃圾分类 扩张卷积 多尺度特征提取 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]