Deep learning-based waste detection in natural and urban environments.

Area of Science:

• Environmental Science
• Computer Vision
• Machine Learning

Background:

• Waste pollution is a significant global environmental issue, necessitating efficient recycling processes.
• Current automatic waste detection research lacks standardized benchmarks and comparable metrics.
• Existing deep learning approaches for waste detection vary widely, hindering progress.

Purpose of the Study:

• To critically analyze existing waste detection datasets and deep learning methodologies.
• To establish a replicable baseline for litter detection through experimental validation.
• To introduce new, unified benchmark datasets (detect-waste and classify-waste) for comprehensive waste categorization.

Main Methods:

• A critical review of over ten waste datasets and existing deep learning waste detection approaches.
• Development of new benchmark datasets by merging and unifying annotations from open-source datasets.
• Implementation of a two-stage detector: EfficientDet-D2 for litter localization and EfficientNet-B2 for classification.
• Semi-supervised training of the classifier using unlabeled images.

Main Results:

• The proposed two-stage detector achieved up to 70% average precision in waste detection.
• The classifier demonstrated approximately 75% accuracy in categorizing detected waste into seven types.
• New benchmark datasets with unified annotations for bio, glass, metal, plastic, non-recyclable, other, paper, and unknown waste were created.

Conclusions:

• The study provides a foundational baseline for reproducible litter detection research.
• The developed datasets and models offer a standardized approach to waste detection and classification.
• Public availability of code and annotations promotes further research and development in automated waste management.