Deep learning in material recovery: Development of method to create training database

Abstract

Increasing the rate of material identification, separation and recovery is a priority in resource management and recovery, and rapid, low cost imaging and interpretation is key. This study uses different combinations of cameras, illuminations and data augmentation techniques to create databases of images to train deep neural networks for the recognition of fibre materials. Using a limited set of 24 material samples sized 1200 cm2, it compares the outcome of reducing them to 30 cm2. The best classification accuracies obtained range from 76.6% to 77.5% indicating it is possible to overcome problems such as limited available materials, time, or storage capabilities, by using a setup with 5 cameras, 5 lights and applying simple software image manipulation techniques. The same method can be used to create deep neural network training databases to recognise a wider range of materials typically found in solid waste streams, in real-time. Furthermore, it offers flexibility as the classification cameras could be deployed at different stages within solid waste processing plants, providing feedback for process control, with the potential of increasing plant efficiency and reducing costs.