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

The global generation of construction, demolition, and renovation (CDR) waste has surged, increasing the demand for efficient recycling solutions. Emerging technologies can automate the sorting of CDR waste, which is crucial for specialised categories such as Electro-construction waste (ECW). However, ECW recovery is particularly challenging due to the heterogeneity and complexity of waste streams. The state-of-the-art segmentation models, trained on generic object images, struggle with part-level recognition tasks, revealing a substantial realm gap. Addressing the gap, this research focuses on automating the recognition of ECW by developing computer vision (CV) models for precise, part-level segmentation of ECW components. The study involves collecting and annotating part-level images to train the model for automatic ECW material recognition and evaluating the performance of several convolutional and transformer backbones. The Swin Transformer outperformed the convolutional neural networks, with a notable improvement of 8.60% and 3.64% over ResNet and ResNeXt, respectively. The study demonstrates the practicality of part-level segmentation for ECW, which is critical for improving resource recovery and promoting sustainable waste management practices.