Nature-Inspired Search Method and Custom Waste Object Detection and Classification Model for Smart Waste Bin
Israel Edem Agbehadji1, Abdultaofeek Abayomi2, Khac-Hoai Nam Bui3
1Honorary Research Associate, Faculty of Accounting and Informatics, Durban University of Technology, P.O. Box 1334, Durban 4000, South Africa.
Sensors (Basel, Switzerland)
|August 26, 2022
Summary
This study introduces a smart waste bin system using the You-Only-Look-Once (YOLO) convolutional neural network for accurate waste classification. The Yolov3 model, combined with a nature-inspired learning rate, demonstrated superior performance in detecting and sorting waste at the point of collection.
Area of Science:
- Computer Science
- Artificial Intelligence
- Environmental Science
Background:
- Global waste management challenges necessitate innovative solutions for effective collection and sorting.
- Existing waste bins lack the capability for at-source solid waste segregation, impacting overall management efficiency.
- The South African University of Technology (SAUoT) is leveraging emerging technologies for advancements in solid waste management.
Purpose of the Study:
- To develop and evaluate a smart waste bin system capable of classifying waste at the point of collection.
- To employ the You-Only-Look-Once (YOLO) object detection algorithm for accurate waste categorization.
- To investigate the efficacy of nature-inspired algorithms for optimizing the learning rate in convolutional neural network (CNN) models for waste detection.
Main Methods:
- A custom YOLO model was developed and trained for waste object detection, utilizing various weights and backbones (darknet53.conv.74, darknet19_448.conv.23, Yolov4.conv.137, Yolov4-tiny.conv.29).
- Eight distinct waste classes were defined, and a dataset of 3171 waste images was used for training and evaluation.
- The Kestrel-based Search Algorithm (KSA) was employed to determine the optimal learning rate for the CNN model.
Main Results:
- The Yolov3 model achieved the highest prediction accuracy with a mean Average Precision (mAP) of 80%, outperforming Yolov3-tiny (57%), Yolov4 (41%), and Yolov4-tiny (74%).
- While Yolov3-tiny offered faster prediction speeds, its accuracy was significantly limited compared to Yolov3.
- The combination of the KSA learning rate (0.0007) and the Yolov3 model proved to be the most accurate for waste object detection and classification.
Conclusions:
- The Yolov3 model, optimized with KSA for learning rate determination, is highly effective for waste object detection and classification in smart waste bin systems.
- Implementing an EdgeIoT-enabled system with Yolov3 can significantly enhance the efficiency of waste collection and management.
- Nature-inspired search algorithms show considerable promise for optimizing parameters in AI models for environmental applications like waste management.
