Construction and demolition waste classification based on vision transformer: Parameter fusion and visualization analysis

Construction and demolition waste (CDW) classification is essential for automated sorting and sustainable material recycling, yet reliable recognition remains challenging due to fine-grained visual similarities among materials and the complex illumination conditions encountered in real working environments. This study proposes a feature fusion framework for CDW image classification that combines deep visual representations with handcrafted descriptors under a unified and reproducible training and evaluation protocol. A labeled dataset containing 29,877 CDW images was constructed, together with four illumination-perturbation test domains, namely underexposed, overexposed, directional-lighting, and mixed-illumination settings. Under the unified configuration, three backbone networks, VGG19, ResNet50, and ViT-B/16, were systematically compared, and additional analyses were conducted on handcrafted feature selection and classification head design. Specifically, nine representative handcrafted features were selected from a 30-feature candidate pool through a learnable gating mechanism, and a classification head based on the Heterogeneous Feature Alignment Module (HFAM) was used to fuse deep and handcrafted features. The results showed that ViT-B/16 exhibited stronger robustness under illumination shifts, especially on the overexposed and mixed-illumination test sets. Meanwhile, the proposed fusion framework provided measurable performance gains, the HFAM design showed good effectiveness and stability, and a more complex explicit modulation strategy did not yield further performance improvement. Overall, this study provides a practical benchmark for CDW classification under realistic illumination variations and presents a reproducible fusion-based pipeline for CDW material identification.