Simultaneous mass estimation and class classification of scrap metals using deep learning

Abstract

While deep learning has helped improve the performance of classification, object detection, and segmentation in recycling, its potential for mass prediction has not yet been explored. Therefore, this study proposes a system for mass prediction with and without feature extraction and selection, including principal component analysis (PCA). These feature extraction methods are evaluated on a combined Cast (C), Wrought (W) and Stainless Steel (SS) image dataset using state-of-the-art machine learning and deep learning algorithms for mass prediction. After that, the best mass prediction framework is combined with a DenseNet classifier, resulting in multiple outputs that perform both object classification and object mass prediction. The proposed architecture consists of a DenseNet neural network for classification and a backpropagation neural network (BPNN) for mass prediction, which uses up to 24 features extracted from depth images. The proposed method obtained 0.82 R2, 0.2 RMSE, and 0.28 MAE for the regression for mass prediction with a classification performance of 95% for the C&W test dataset using the DenseNet+BPNN+PCA model. The DenseNet+BPNN+None model without the selected feature (None) used for the CW&SS test data had a lower performance for both classification of 80% and the regression (0.71 R2, 0.31 RMSE, and 0.32 MAE). The presented method has the potential to improve the monitoring of the mass composition of waste streams and to optimize robotic and pneumatic sorting systems by providing a better understanding of the physical properties of the objects being sorted.