Honorary Research Associate, Faculty of Accounting and Informatics, Durban University of Technology, P.O. Box 1334, Durban 4000, South Africa.
Department of Information and Communication Technology, Mangosuthu University of Technology, P.O. Box 12363, Durban 4026, South Africa.
Sensors (Basel). 2022 Aug 18;22(16):6176. doi: 10.3390/s22166176.
Waste management is one of the challenges facing countries globally, leading to the need for innovative ways to design and operationalize smart waste bins for effective waste collection and management. The inability of extant waste bins to facilitate sorting of solid waste at the point of collection and the attendant impact on waste management process is the motivation for this study. The South African University of Technology (SAUoT) is used as a case study because solid waste management is an aspect where SAUoT is exerting an impact by leveraging emerging technologies. In this article, a convolutional neural network (CNN) based model called You-Only-Look-Once (YOLO) is employed as the object detection algorithm to facilitate the classification of waste according to various categories at the point of waste collection. Additionally, a nature-inspired search method is used as learning rate for the CNN model. The custom YOLO model was developed for waste object detection, trained with different weights and backbones, namely darknet53.conv.74, darknet19_448.conv.23, Yolov4.conv.137 and Yolov4-tiny.conv.29, respectively, for Yolov3, Yolov3-tiny, Yolov4 and Yolov4-tiny models. Eight (8) classes of waste and a total of 3171 waste images are used. The performance of YOLO models is considered in terms of accuracy of prediction (Average Precision-AP) and speed of prediction measured in milliseconds. A lower loss value out of a percentage shows a higher performance of prediction and a lower value on speed of prediction. The results of the experiment show that Yolov3 has better accuracy of prediction as compared with Yolov3-tiny, Yolov4 and Yolov4-tiny. Although the Yolov3-tiny is quick at predicting waste objects, the accuracy of its prediction is limited. The mean AP (%) for each trained version of YOLO models is Yolov3 (80%), Yolov4-tiny (74%), Yolov3-tiny (57%) and Yolov4 (41%). This result of mAP (%) indicates that the Yolov3 model produces the best performance results (80%). In this regard, it is useful to implement a model that ensures accurate prediction to develop a smart waste bin system at the institution. The experimental results show the combination of KSA learning rate parameter of 0.0007 and Yolov3 is identified as the accurate model for waste object detection and classification. The use of nature-inspired search methods, such as the Kestrel-based Search Algorithm (KSA), has shown future prospect in terms of learning rate parameter determination in waste object detection and classification. Consequently, it is imperative for an EdgeIoT-enabled system to be equipped with Yolov3 for waste object detection and classification, thereby facilitating effective waste collection.
废物管理是全球各国面临的挑战之一，因此需要创新的方法来设计和操作智能垃圾桶，以实现有效的废物收集和管理。现有的垃圾桶无法在收集点促进固体废物的分类，这对废物管理过程产生了影响，这就是这项研究的动机。南非开普敦理工大学（SAUoT）被用作案例研究，因为固体废物管理是 SAUoT 通过利用新兴技术产生影响的一个方面。在本文中，使用了一种称为单次检测（YOLO）的卷积神经网络（CNN）模型作为目标检测算法，以便在废物收集点根据各种类别对废物进行分类。此外，还使用了一种受自然启发的搜索方法作为 CNN 模型的学习率。定制的 YOLO 模型是为废物目标检测开发的，使用不同的权重和骨干网络进行训练，分别为 darknet53.conv.74、darknet19_448.conv.23、Yolov4.conv.137 和 Yolov4-tiny.conv.29，用于训练 Yolov3、Yolov3-tiny、Yolov4 和 Yolov4-tiny 模型。使用了 8 类废物和总共 3171 张废物图像。YOLO 模型的性能是根据预测的准确性（平均精度-AP）和以毫秒为单位测量的预测速度来考虑的。损失值的百分比越低表示预测性能越高，预测速度越低。实验结果表明，Yolov3 的预测精度优于 Yolov3-tiny、Yolov4 和 Yolov4-tiny。虽然 Yolov3-tiny 预测废物对象的速度很快，但它的预测精度有限。每个经过训练的 YOLO 模型的平均 AP（%）分别为 Yolov3（80%）、Yolov4-tiny（74%）、Yolov3-tiny（57%）和 Yolov4（41%）。mAP（%）的这一结果表明，Yolov3 模型产生的性能结果最好（80%）。在这方面，在机构中实施一个确保准确预测的模型来开发智能垃圾桶系统是有用的。实验结果表明，KSA 学习率参数为 0.0007 与 Yolov3 的结合被确定为废物目标检测和分类的准确模型。在废物目标检测和分类方面，使用受自然启发的搜索方法（如基于游隼的搜索算法（KSA））在确定学习率参数方面显示出了未来的前景。因此，一个支持 EdgeIoT 的系统需要配备 Yolov3 来进行废物目标检测和分类，从而有效地进行废物收集。
