Determination of the composition of recycled aggregates using a deep learning-based image analysis

Abstract

Recycled aggregates (RA) are obtained by crushing inert construction and demolition waste. Their composition is variable and is currently obtained by a time-consuming manual sorting. Our work makes use of deep learning, especially convolutional neural networks (CNN), to determine this composition in near real time and in an automated way. A labelled database was created for learning of the CNNs. It consists of approximately 36,000 images of individual grains classified according to their nature. After training, our best-performing CNN reaches a validation accuracy of 97% for classifying images of grains. It is based on a Residual Network that we customised in order to improve its performance. Moreover, we evaluated the mass of the grains by assuming that grains of a given nature have a constant form and density. Our approach was compared with manual sorting. There was less than 2% of difference in mass for most RA natures tested.