Automated recognition of contaminated construction and demolition wood waste using deep learning

Abstract

Wood waste is a significant component of construction and demolition waste; however, contamination often limits its recovery. Efficient sorting of contaminated wood remains underexplored, with current methods relying heavily on manual separation. This study proposes a deep learning-based approach to classify wood waste by contamination type using RGB images, addressing a critical gap in the resource recovery context. A custom dataset of six common contaminated wood waste types was curated, and four selective advanced deep learning models, including convolutional neural networks (RegNet, ConvNeXt) and Transformers (Vision Transformer, Swin Transformer), were evaluated using transfer learning. Remarkably, ConvNeXt outperformed all models with 91.67 % validation accuracy and balanced performance, with precision, recall, and F1 score around 0.8667. The results highlight deep learning's potential to enhance construction and demolition wood waste management, enabling automated sorting and supporting reusing, recycling, and reclaiming efforts, thereby reducing landfill dependency and promoting sustainable practices.