Classification of Recycled Aggregates Using Deep Learning

Abstract

AbstractThe European Union is promoting sustainable development and circular economy by inciting its member states to recycle at least 70% of their construction and demolition waste (CDW) through the Horizon 2020 programme. CDW is crushed in order to obtain recycled aggregates (RA). The latter are a mixture of recycled concrete aggregates, natural stones, clay bricks, bituminous grains and with other materials (e.g. glass, wood and steel). The composition of RA is variable and in order to determine it, the NF EN 933-11 standard recommends manual sorting. However, it is time-consuming and it is performed only occasionally on recycling plants. Our work focuses on the development of a novel method to determine the composition of RA faster and in an automated way. It makes use of deep learning, particularly convolutional neural networks (CNN). CNNs can analyse images of RA and identify the nature of every aggregate in order to give the composition of the RA instantly. 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. The best-performing CNN is now able to identify correctly the class (i.e. the nature) of 97% of aggregates present on a picture of RA. Moreover, we proposed a method to evaluate the mass of the grains by assuming that those of a given nature have a constant form and density. We are also working on the automatic extraction of the grains from a picture of RA using Mask R-CNN.KeywordsCircular economyConvolutional neural networksDeep learningImage analysisRecycled aggregates