Sorting of scrap metal using self-supervised learning and transformers
||Sorting of scrap metal using self-supervised learning and transformers|
||Felipe Boeira <email@example.com>|
||2022-02-16 – 2022-09-01|
Recycling metal requires that it is sorted by material, but current methods lack automatic validation of how well the sorting is performed. Recycling aluminium contributes to reducing climate emissions as recycling 1 ton of aluminium reduces carbon dioxide emissions by 13 tons. The Stena recycling plant in Halmstad handles 300 tons of refuse every day which requires a large amount of manual labour to sort. Hence, it is essential that the material is properly sorted before it is melted, otherwise, the resulting metal might not be usable. Earlier work has explored using machine learning techniques to perform automatic validation of the sorting using a convolutional neural network on images of scrap metal. While previous methods have required labour-intensive manual annotation of training images, this project aims to investigate self-supervised learning to mitigate the amount of manual annotation needed. In addition to employing self-supervised learning, we investigate how transformers can be used as an alternative technique to the convolutional neural networks used in prior works.