Optimizing construction and demolition waste Segmentation using Advanced Deep Learning Techniques (Semantic vs Instance segmentation)

Rapid urbanization has accelerated the generation of C&D wastes in large quantities, therefore, leading to environmental as well as economic issues. The conventional methods of segregating wastes by hand and evaluating volumes are inefficient and susceptible to errors; therefore, an automated method is desired. This paper evaluates whether deep learning-based segmentation models can be used to study C&D waste, primarily focusing on concrete, tiles, and bricks. The two models that are used are a U-Net-based model for semantic segmentation and Mask R-CNN for instance segmentation. Mask R-CNN gave better performance with better accuracy in segmentation and was also consistent up to 5500 annotated images and provided better results in detailed irregular shapes. The U-Net-based model performed relatively well in semantic segmentation but failed in the case of complex structures where instance-level differentiation is required. Both models were further optimized with Optuna in hyper parameters, which allowed increasing their accuracy. In summary, these results show Mask R-CNN to be a more effective tool for detailed waste segmentation, leading to significant implications for improvement of recycling processes and reductions of environmental impacts.