WMC-ViT: Waste Multi-class Classification Using a Modified Vision Transformer
Aidan G. Kurz, Ethan R. Adams, Arthur C. Depoian, Colleen P. Bailey, Parthasarathy Guturu · 2022
The constant production and lack of efficient waste management procedure has created a need for automated classification of trash as it comes into facilities. This paper proposes a new algorithm for efficiently classifying objects found in solid waste processing by utilizing a combination of vision transformers (ViT) and convolutional neural networks (CNNs) to create a Multi-Head block for parallel processing of multiple transformers. This method identifies five unique classes of the most common material found in waste with peak test accuracy of 94.27% using 35492 total parameters, a reduction of 99.74% when compared to current state of the art methods, allowing for lower power operations and easier deployment.