Self-prompting segmentation framework: an advanced computer vision technique for electro-construction waste valorisation

The effective recognition and sorting of Electro-Construction Waste (ECW) remain critical challenges in Material Recovery Facilities (MRFs), limiting resource valorisation and compromising worker safety. Computer vision shows strong potential for automated waste recognition and recycling. However, its application to recognising ECW components remains unexplored. This study introduces the "Self-Prompting Segmentation Framework" (SPSF), a parameter-efficient computer vision pipeline for precise part-level segmentation of ECW, enabling automated recognition, dismantling, and sorting in MRFs. The framework adapts a large-scale vision foundation model by integrating a detector for self-prompt generation, and Low-Rank Adaptation for efficient fine-tuning, updating only a small subset of parameters while preserving high segmentation fidelity. Validated on a curated dataset of real-world ECW images, SPSF achieves relative improvements of 39.17 % in mean Intersection over Union, 39.27 % in mean Dice-Sørensen Coefficient, and 28.15 % in mean Average Precision over the Segment Anything Model 2 baseline. These results demonstrate that SPSF can accurately identify individual components within different ECW categories. This capability provides a practical and scalable approach for automated waste component recognition, supporting more efficient material sorting and recycling.