Waste Object Detection and Classification using Deep Learning Algorithm: YOLOv4 and YOLOv4-tiny

Abstract

As industrialization, urbanization, and global population levels increased, so did the pollution and environmental
degradation because of the waste. The common problem related to waste is the waste sorting that is still improperly handled
from the household level to the final disposal site. The solution offered in this research is using deep learning algorithm for
object detection using YOLOv4 and YOLOv4-tiny with Darknet-53. The dataset consists of 3870 waste images divided into
4 classes like glass, metal, paper, and plastic. At the testing stage, each model uses 3 different inputs such as images, videos,
and webcams. In the YOLOv4-tiny model, experiments hyperparameter were also carried out on subdivision values and
mosaic data augmentation. The result proves that YOLOv4 have better performance than YOLOv4-tiny for object detection,
although in terms of computational speed the YOLOv4-tiny’s scores are better. The best results from the YOLOv4 model
reach mAP 89.59%, precision 0.76, recall 0.90, F1-score 0.82, and Average IoU 64.01%, while the best YOLOv4-tiny results
are mAP 81, 84%, precision 0.59, recall 0.83, F1-score 0.69, and Average IoU 48.35%. This research also proves that the
models with smaller subdivision values and uses a mosaic have an optimal performance.