Litter Detection with Deep Learning: A Comparative Study.

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Litter pollution poses a significant environmental challenge.
• Automated litter detection aids in waste assessment and supports environmental monitoring.
• Existing research lacks focus on deep learning for litter detection on low-power devices.

Purpose of the Study:

• To evaluate state-of-the-art deep learning object detection models for litter detection.
• To assess model performance on resource-constrained devices like smartphones.
• To introduce a new dataset for litter detection research.

Main Methods:

• Comparative analysis of Convolutional Neural Network (CNN) architectures (Faster RCNN, Mask-RCNN, EfficientDet, RetinaNet, YOLO-v5).
• Utilized two existing litter image datasets and introduced the new PlastOPol dataset (2418 images, 5300 annotations).
• Evaluated models on a smartphone to simulate real-world, low-processing capability scenarios.

Main Results:

• YOLO family object detectors demonstrated superior performance in litter detection.
• YOLO models achieved high accuracy, fast processing times, and low memory footprint.
• These findings indicate suitability for deployment on mobile devices.

Conclusions:

• Deep learning, particularly YOLO models, offers a viable solution for automated litter detection.
• The PlastOPol dataset enhances resources for environmental waste research.
• This work paves the way for efficient, mobile-based environmental monitoring tools.