Real-time instance segmentation of recyclables from highly cluttered construction and demolition waste streams

Area of Science:

• Environmental Science
• Computer Science
• Robotics

Background:

• Escalating construction and demolition waste (CDW) necessitates efficient recycling strategies.
• Automated identification of CDW recyclables using deep learning is promising but challenged by clutter and complexity.
• Accurate and fast localization is essential for robotic waste sorting.

Purpose of the Study:

• To comprehensively assess state-of-the-art real-time instance segmentation for recyclables in complex CDW streams.
• To address the limitations of current deep learning models in handling CDW intricacies like deformation, contamination, and clutter.
• To propose and evaluate novel techniques for improving the accuracy and efficiency of CDW recyclable localization.

Main Methods:

• Curated and employed a high-quality CDW instance segmentation dataset capturing real-world complexities.
• Assessed state-of-the-art deep learning networks for real-time instance segmentation.
• Integrated patch-based inferencing techniques to enhance focus on cluttered regions.
• Developed a domain-transfer framework to improve zero-shot capabilities of prompt-based segmentation.

Main Results:

• Advanced networks achieved less than 50% segmentation accuracy on complex CDW streams due to high clutter.
• Patch-based inferencing boosted overall performance by 12.9%.
• The proposed domain-transfer framework increased correct mask predictions by 68% for zero-shot identification.

Conclusions:

• Current deep learning models face significant challenges in segmenting recyclables from complex CDW.
• Patch-based inferencing and domain-transfer frameworks offer practical solutions to enhance automated CDW recycling.
• This study provides a reference for applying deep learning in environmental management and suggests optimal architectural frameworks for waste sorting.