作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (8): 294-300. doi: 10.19678/j.issn.1000-3428.0058258

• 开发研究与工程应用 • 上一篇    下一篇

基于改进Faster R-CNN的垃圾检测与分类方法

马雯1, 于炯1,2, 王潇1, 陈嘉颖2   

  1. 1. 新疆大学 软件学院, 乌鲁木齐 830008;
    2. 新疆大学 信息科学与工程学院, 乌鲁木齐 830046
  • 收稿日期:2020-05-07 修回日期:2020-07-02 发布日期:2020-07-23
  • 作者简介:马雯(1993-),女,硕士研究生,主研方向为深度学习、图像识别;于炯(通信作者),教授、博士生导师;王潇,硕士研究生;陈嘉颖,博士研究生。
  • 基金资助:
    国家自然科学基金(61862060,61462079,61562086,61562078)。

Garbage Detection and Classification Method Based on Improved Faster R-CNN

MA Wen1, YU Jiong1,2, WANG Xiao1, CHEN Jiaying2   

  1. 1. College of Software, Xinjiang University, Urumqi 830008, China;
    2. College of Information Science and Engineering, Xinjiang University, Urumqi 830046, China
  • Received:2020-05-07 Revised:2020-07-02 Published:2020-07-23

摘要: 针对人工分拣垃圾环境差、任务繁重且分拣效率低的问题,为提高垃圾识别与分类的精确度,同时克服垃圾体积小及图像分辨率较低的难题,基于现有深度卷积神经网络模型,提出改进的Faster R-CNN目标检测模型与VGG16及ResNet50卷积神经网络相结合的方法。根据卷积网络的特性,修改Faster R-CNN网络结构,提升小目标检测任务精度,采用Soft-NMS算法替代传统的非极大值抑制算法,并对参数进行敏感分析,确定其参数范围为0.4~0.7。实验结果表明,与传统Faster R-CNN算法相比,该方法平均精确度提高8.26个百分点,综合识别率达到81.77%,且能够减少图像处理时间。

关键词: 垃圾识别, 分类, 目标检测, 深度学习, 非极大值抑制

Abstract: Traditional garbage sorting is implemented manually. Workers have to suffer from the poor environment and intensive tasks, which reduce their efficiency. Neural network can help automate the process, and thus free workers from the heavy labor. This paper proposes a neural network-based method to improve the accuracy of garbage identification and classification, especially for the small-sized garbage and low-resolution images. The method combines the improved Faster R-CNN target detection model, VGG16 and Resnet50. According to the features of the convolutional network, the structure of Faster R-CNN is modified to improve the accuracy of small target detection. The Soft-NMS algorithm is used to replace the traditional non-maximum suppression algorithm, and conduct sensitive analysis of the parameters to determine the parameter ranging within 0.4~0.7. Experimental results show that compared with the traditional Faster R-CNN algorithm, the method increases the average accuracy by 8.26 percentage points, and provides a comprehensive recognition rate of up to 81.77%. It also reduces the image processing time.

Key words: garbage identification, classification, target detection, deep learning, non-maximum suppression

中图分类号: