SKM 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
HL: Fachverband Halbleiterphysik
HL 51: Nitrides: Preparation and Characterization
HL 51.2: Vortrag
Freitag, 31. März 2023, 09:45–10:00, POT 112
yolo-assisted object detection of dislocation-related pits on GaN surfaces using generative adversarial networks — •Mahdi Khalili Hezarjaribi1,2, Uwe Rossow1,2, Heiko Bremers1,2, and Andreas Hangleiter1,2 — 1Institute of Applied Physics, Technische Universität Braunschweig, Germany — 2Laboratory for Emerging Nanometrology, Technische Universität Braunschweig, Germany
In this paper, we present a model for detecting dislocation-related pits on GaN surfaces using SEM images, which strongly relies on the use of synthetic image generation. Pits mark dislocations in the layers and are widely used to assess crystalline quality; therefore, a considerable amount of images containing pits have to be evaluated. For this purpose, a Deep Learning (DL) algorithm is employed to achieve objective results, which are detecting the pits, and hence the dislocations, in SEM images. In order to train the algorithm to efficiently detect the objects, we need a host of SEM images containing pits in multiple sizes, numbers, formations, and noises. Due to the complexity of the microscopic structures and the lack of enough images, we incorporated a group of powerful algorithms called Generative Adversarial Networks to create artificial fake images just like real images and feed them, together with our real images, to our dataset to enrich the volume of the dataset. In the next stage, the YOLO algorithm (version 5) has been employed as the core deep learning algorithm for the object detection process using the above-mentioned dataset to train the network. A minimum average confidence of 86% for detecting real objects has been realized, corresponding to a high probability of detection.