Fine metal-rich waste stream characterizationbased on RGB data: Comparison betweenfeature-based and deep learning classificationmethods

Background: Material compositions in the recycling industry
are currently determined by manual sorting, which is
time intensive and shows subjective influences. For an automated,
sensor-based material flow characterization a particlebased
material classification is necessary. Aim: The classification
of metal-containing fine-fractions based on RGB images with
different machine learning (ML) techniques is investigated on
two created datasets A (12,480 images) and B (19,498 images).
Method: Two approaches are compared: In approach I, images
are firstly pixel- and then object-based classified with six different
ML models on three color spaces. In approach II, images are
classified by six different convolutional neural networks (CNNs).
Results: The classification of dataset A was possible with high
accuracy (> 99.8 %) for both approaches and chosen ML algorithms
were of minor importance. For dataset B, approach I
achieved an accuracy of 78.2 % ± 2.0 %, and chosen ML algorithms
were of higher importance for object-based classification.
In approach II, the best-performing CNN achieved an accuracy
of 80.4 % ± 4.2 % and a top-3 score of 94.2 % ± 2.6 %. Conclusion:
Results from existing studies for coarser particle sizes can
be transferred to fine fractions. Further research is needed to improve
the classification of dataset B, e. g. by adding instances to
less frequent classes and applying deeper CNNs.