Automated Electro-construction waste Sorting: Computer vision for part-level segmentation

Area of Science:

• Environmental Science
• Computer Science
• Materials Science

Background:

• Global construction, demolition, and renovation (CDR) waste generation is increasing, necessitating advanced recycling solutions.
• Electro-construction waste (ECW) recovery is complex due to heterogeneous waste streams, posing challenges for automated sorting.
• Current segmentation models lack part-level recognition capabilities for ECW, creating a significant performance gap.

Purpose of the Study:

• To develop and evaluate computer vision (CV) models for automated, part-level segmentation and recognition of ECW components.
• To address the limitations of existing models in handling the complexity of ECW.
• To enhance the efficiency of ECW material recognition for improved resource recovery.

Main Methods:

• Collection and annotation of part-level images specific to ECW.
• Development of CV models, including convolutional neural networks (CNNs) and transformer backbones, for ECW segmentation.
• Evaluation of model performance using metrics for precise, part-level recognition.

Main Results:

• The Swin Transformer model demonstrated superior performance in part-level ECW segmentation compared to CNNs.
• Swin Transformer achieved an 8.60% improvement over ResNet and a 3.64% improvement over ResNeXt.
• The study confirmed the feasibility of part-level segmentation for accurate ECW material recognition.

Conclusions:

• Part-level segmentation using advanced CV models like the Swin Transformer is practical and effective for ECW recognition.
• Improved ECW recognition is critical for enhancing resource recovery and advancing sustainable waste management practices.
• This research contributes to automating the recycling of complex waste streams, promoting a circular economy.