Autonomous textile sorting using hyperspectral imaging

Textile sorting for recycling is a challenging task, currently performed manually by trained operators relying on cloth tags and on their own knowledge - a method that is highly expensive and time-consuming, and potentially unreliable. This research reports the results of material classification of fabrics using a Hyperspectral camera in the Visible Near Infrared Range (VNIR), which is a more economically viable sensor than the NIR sensor, which currently dominates research in this area. We compare the results of two methodologies that were used to classify the data, a Shallow Neural Network (NN) algorithm and a Convolutional Neural Network (CNN). Results show that NNs can quickly recognise pure materials, but difficulties arise with blended materials. CNNs are most effective in identifying small non-fabric features like buttons and zips. However, a wide range of samples and methodologies would be needed before establishing a viable, scalable system.