AI-baserad identifiering av kopparrika komponenter i metallskrot för robotiserad sortering : En jämförelse mellan bounding box-detektion och segmentering

Sammanfattning

In end-of-life vehicle recycling, copper is a critical contaminant in recovered steel, where even low concentrations cause hot shortness and degrade the mechanical properties of the material. When vehicles are fragmented, copper-rich components are mixed into the steel scrap stream. Conventional separation methods such as magnetic sorting and eddy-current separation cannot fully distinguish copper from other non-ferrous fractions. This creates a need for automated vision systems capable of identifying copper-rich objects before fragmentation.The purpose of this thesis is to investigate how an image-based AI system can be used to identify copper-rich components in a mixed metal scrap stream, as a first step towards robotic sorting of such components, and to analyse the opportunities and limitations of the application. The study focuses on two component types that typically contain copper, cables and printed circuit boards, and compares two representation methods: bounding box detection (YOLOv8) and instance segmentation (YOLOv8-seg).The work was conducted as an experimental study within the Mission 0 House initiative at Uppsala University in collaboration with Polestar. Image data was collected in a controlled laboratory environment using an Intel RealSense D435 camera, annotated in Roboflow with SAM3 assistance, and split into training, validation, and test sets. The models were evaluated using precision, recall, and mAP, with recall as the primary metric since missed copper objects lead to irreversible material contamination. During the work, annotation was identified as the main practical bottleneck of the project, in particular for the instance segmentation track. As an exploratory side-track, an automated annotation pipeline combining a vision-language model with SAM2 was therefore investigated. To verify end-to-end integration with physical hardware, the final segmentation model was deployed on an NVIDIA Jetson AGX Thor connected to a UR5 robot in a small proof-of-concept setup.Initial results on a limited dataset showed that the bounding box model achieved mAP@0.5 = 0.91, precision = 0.91 and recall = 0.88 when trained on 104 images, while the segmentation model trained on 120 images performed considerably worse (mAP@0.5 = 0.67, recall = 0.60). In a second iteration, where the segmentation dataset was expanded to approximately 1,750 images, the model improved substantially in terms of mask stability, generalisation to thin cables and confidence during live inference. This illustrates the high data requirements of instance segmentation rather than an inherent unsuitability of the method.