OSITY: A Deep Learning and Computer Vision Framework for Tiny Wood Particle Detection, Classification, and Sorting

The increasing global output of end-of-life fibreboards provides needs for highly efficient and precise separation techniques to support tertiary recycling and valorization processes. Current industrial practice relies on multi-modal sensing often combined with manual inputs, a process susceptible to human factors such as fatigue, leading to compromised throughput and accuracy trade-offs. This study introduces the Optical Segmentation and Identification Tracking with YOLO (OSITY) Model, a deep learning architecture that was customized for real-time sorting with speed and accuracy, fine-grained instance segmentation and classification of post-consumer wood and non-wood particles with sizes about 8 mm. The dataset contained 5 consolidated classes described as Solid Wood, Wood Based Panels, Fibreboard Pure, Fibreboard Coating, and Non-wood. The results showed promising application on correctly segmenting and identifying the tiny particles with a validation accuracy of 83%. This result shows that despite similar physical features of wood particles like Fibreboard with and without coating, several features could be extracted and predicted. Such results demonstrate the use of advanced techniques in wood recycling.