Deep learning computer vision for the separation of Cast- and Wrought-Aluminum scrap

Abstract

In consequence of the electrification and the increased adoption of lightweight structures in the automotive industry, global demand for wrought Aluminum (Al) is expected to rise while demand for cast Al will stagnate. Since cast alloys can only be converted to wrought alloys by energy-intensive processes, the most promising strategy to avoid the emergence of excess Al cast alloys scrap is to sort cast from wrought alloys. To date, the separation of complex mixes of non-ferrous metals often implies the use of either or both sink-float techniques and/or X-ray fluorescence (XRF) based sorting. Therefore, the presented research develops an efficient method to classify cast and wrought (C&W) alloys in a real-time system with a conveyor belt using transfer learning methods, such as fine-tuning and feature extraction. Five CNNs are evaluated to classify C&W alloys using colour and depth images and transfer learning methods. In addition, the early fusion and late fusion of colour and depth images of C&W Al are investigated. For early fusion, data is added as an extra input channel to the first convolution layer of the CNN, and for later fusion, the images are fed in two separate subnetworks with the same architecture, where the parameters of the fully-connected layers are concatenated in both subnetworks. Our approach shows that late fusion CNN DenseNet allows obtaining the best performances and can achieve up to 98% accuracy.