Outdoor Trash Detection in Natural Environment Using a Deep Learning Model

• Published in: IEEE access 2023-01, Vol.11, p.1-1
• Main Authors: Das, Dhrubajyoti, Deb, Kaushik, Sayeed, Taufique, Dhar, Pranab Kumar, Shimamura, Tetsuya
• Format: Article
• Language: English
• Subjects: Accuracy; Bangladeshi trash; Benchmark testing; Classification; Data collection; Datasets; Deep learning; Detectors; Localization; Machine learning; Object detection; Object recognition; single-stage object detector; Task analysis; Transfer learning; trash dataset; trash detection; Waste management
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3313166

Trash production and disposal have emerged as serious issues for underdeveloped nations as their populations have swelled. As manual classification can be both time-consuming and potentially dangerous, therefore, nowadays, it is increasingly being replaced by automated methods. Recent advances in AI and deep learning have allowed for significant advancements in trash detection and classification systems. Due to the lack of a comprehensive trash detection dataset tailored to Bangladesh, we set out to collect data that would accurately portray the complexity of Bangladesh's scenario while also incorporating openlittermap. In this study, we employ a deep learning model known as YOLOv5. Several variants of the YOLOv5 model are used and assessed with both the freshly minted dataset and the already existing benchmark datasets. Simulation results indicate that the finetuned YOLOv5 model outperforms existing models in terms of mean average precision (mAP) and F1-score. On the Bangladeshi dataset, the model shows an mAP of 34.3% and an F1-score of 43.7%. The mAP and F1-score provide a holistic evaluation of YOLOv5's object recognition accuracy, localization, and precision-recall balance. By incorporating the additional data from openlittermap into the new dataset, the mAP is increased to 45.4%. In addition, for some variants of YOLOv5, the suggested model produces greater mAP than the current literature on both the TACO and PlastOpol datasets. The model also achieves an mAP of 84.4% and an F1-score of 78.2% in single-class detection experiments with the newly created dataset. This is because concentrating on just one class helps eliminate class ambiguity, improves localization accuracy, and mitigates class imbalance.