Robust referring image segmentation for construction and demolition waste recognition.

Automated sorting is critical for improving the recycling rates of Construction and Demolition Waste (CDW), a significant global waste stream. While language-guided recognition-a technology known as Referring Image Segmentation (RIS)-offers a potential frontier for flexible and intelligent sorting, its practical deployment is hindered by a considerable lack of robustness. These systems often fail when operator instructions are ambiguous or do not correspond to objects in the visual scene-common occurrences in dynamic industrial environments-leading to sorting errors and reduced material purity. To address this challenge, this study introduces RefSegformer-CDW, a robust recognition framework explicitly designed to enhance the reliability of automated CDW sorting. The architecture integrates a cross-modal verification mechanism that enables the system to reject invalid commands, preventing costly sorting errors. The Ref-CODD dataset was developed to facilitate the development and validation of such robust systems, the first benchmark featuring real-world CDW imagery paired with adversarial negative samples that simulate challenging operational scenarios. Experimental results on the Ref-CODD dataset demonstrate the effectiveness of the proposed RefSegformer-CDW. It achieves a mean Intersection over Union (mIoU) of 94.03%, surpassing the state-of-the-art by 1.24 percentage points. The framework shows high stability under varying illumination, with the mIoU dropping by only 2.38 percentage points between dark and normal lighting conditions. Moreover, it maintains robust performance across diverse waste types, achieving an average overall Intersection over Union (oIoU) of 93.90% on ten different CDW categories.