| 摘 要: 为解决垃圾分类收运过程中，由于错误收运导致的混装率高、垃圾分类效果下降的问题，文章提出了一种基于改进Resnet18的垃圾分类收运监管方法。首先，在Resnet18残差结构的始端引入SE-Net通道注意力机制模块，可以有效提升网络的特征提取能力；其次，采用一种基于空洞卷积的多尺度感受野融合模块，使网络能够融合不同尺度的特征信息。实验结果表明，残差结构始端和晚期融合的改进方案效果更佳，加权F1值较原Resnet18分别提升了9.26%和7.36%。改进后的模型加权的F1值达97.27%，较原Resnet18分别提升了10.26%，并且对不同颜色垃圾桶识别的准确率更高。此外，改进后的Resnet18的每秒检测帧数(FPS)达到65.03，可以满足监管实时性的要求。同时，采用数据增强和天气模拟的方法处理数据集，使模型能适应多种环境，提升了模型的鲁棒性。 |
| 关键词: Resnet18；垃圾分类收运监管；SE-Net；多尺度感受野融合；扩张卷积 |
|
中图分类号: TP391.4
文献标识码: A
|
|
| Waste Classified Collection and Transportation Supervision Approach based on Improved Resnet18 |
|
HE Yanhong1, XU Yining2, FU Jiaqi1, CHEN Shuhang1, LI Junfeng1
|
( 1.School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China ;
2.School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China)
2019330301193@mails.zstu.edu.cn; 2019329621218@mails.zstu.edu.cn; 2019330301190@mails.zstu.edu.cn; 2019330361008@mails.zstu.edu.cn; ljf2003@zstu.edu.cn
|
| Abstract: In order to solve the problem of high mixed loading rate and poor waste classification effect caused by incorrect collection and transportation in the process of waste classification and transportation, this paper proposes a waste classification and transportation supervision approach based on improved Resnet18. Firstly, SE-Net (Squeeze-and-Excitation Networks) channel attention mechanism module is introduced to the beginning part of residue structure of the Resnet18, which can effectively improve the network feature extraction capability. Secondly, a multi-scale receptive field fusion module based on dilation convolution is adopted to enable the network to feature information of different scales. The experimental results show that the improved model of infusion of the beginning part and the late part of the residual structure performs better, and the weighted F1 value increases by 9.26% and 7.36% respectively compared with the original Resnet18. The weighted F1 value of the improved model reaches 97.27%, which is 10.26% higher than that of the original Resnet18, and the recognition accuracy of different color trash cans is higher. In addition, the number of frames per second (FPS) of the improved Resnet18 reaches 65.03, which can meet the requirements of real-time monitoring. At the same time, data argumentation and weather simulation methods are used to process the dataset, so that the proposed model can adapt to a variety of environment and improve the robustness of the model. |
| Keywords: Resnet18; waste classification and transportation supervision; SE-Net; multi-scale receptive field fusion; dilated convolution |
