Implementation and evaluation of a real-time capable approach to sensor-based sorting using CNNs

In state-of-the-art optical sorting, engineered image processing algorithms are commonly used to identify materials. However, for complex material textures and high-density material streams, these approaches are often unable to maintain the desired sorting quality. Convolutional neural networks (CNNs) have been proven to outperform such approaches in their ability to classify objects in images. Nevertheless, most studies are not implemented into a real sorting system. In this work, we implement a full pipeline for optical sorting based on a CNN. Semantic segmentation is performed on acquired images and the result is mapped onto pneumatic valves for ejection of detected objects. The approach is real-time capable, even if the images are captured after discharge from the belt. The effectiveness of the approach is demonstrated in two challenging sorting tasks: the sorting of construction and demolition waste with a high amount of dust and the detection of peanuts in hibiscus tea.