A Deep Learning Approach for Real-Time Segmentation of Graphene Layers Using the YOLO11-Seg Architecture

Abstract

Automated segmentation of graphene layers is a crucial step in the characterization of 2D materials, demanding high precision to identify variations in layer thickness. This study develops an automated system based on the You Only Look Once version 11–segmentation (YOLO11-seg) architecture to accurately detect and segment graphene layers. The dataset used in this work comprises 1,775 optical microscope images classified into four thickness categories: 1-Layer, 2-Layer, 3-Layer, and 4-Layer. The YOLO11-seg model was trained and evaluated three times under identical configurations to assess its performance consistency. The evaluation results from the three training runs demonstrate that the YOLO11-seg model achieved consistent and high performance. The model attained overall precision (P) and recall values of 0.66 and 0.67, respectively. The mean Average Precision (mAP) at an IoU threshold of 0.50 (mAP50) reached 0.71, while the mAP at IoU 0.50-0.95 (mAP50-95) was 0.45. These metrics, combined with an inference speed of 51.5 FPS, indicate high model convergence, stability, and efficiency suitable for real-time applications. For practical implementation, the system was deployed as a web-based application featuring a Node.js interface, with the inference process managed by a Flask API to facilitate real-time segmentation. The results of this research highlight the strong potential of the YOLO11-seg model for rapid and accurate analysis of 2D materials, supporting advancements in nanomaterial research and industrial applications.