Detection of Polymer Thickness Based on the Degree of Their Transparency in SWIR Spectrum on Edge Computing Device

Abstract

The study demonstrates that polymer (ABS) parts and their thickness can be accurately detected using SWIR camera images. We also evaluated the feasibility of performing this detection on edge computing devices like the Jetson Nano, ensuring sufficient speed for real-time applications. For this purpose, we used models from the latest versions of the YOLO object detection framework. Using a dataset of 870 images, featuring circular samples with thicknesses ranging from 0.2 to 1 mm, we trained and tested the neural networks. The YOLOv9-E model achieved the highest precision with an AP@[.5:.05:.95] of 96.57%. In particular, even smaller, faster models like the YOLOv8s demonstrated high precision (AP@[.5:.05:.95] of 96.28%) and rapid processing times, with an average inference time of 70.6 ms. These results suggest that SWIR imaging combined with YOLO models is a viable approach for real-time industrial applications, particularly in the recycling of polymer waste scenarios.