Deep Learning-Based Segmentation of Nanomaterial Images Acquired via Scanning Electron Microscopy

• Ivânia Maria Lucinda De Donato UFRJ http://orcid.org/0009-0008-7639-6423
• Juliana do Nascimento Lunz INMETRO https://orcid.org/0000-0002-0341-5418
• Fernanda Davi Marques INMETRO https://orcid.org/0000-0003-2652-557X
• Bráulio Soares Archanjo INMETRO
• Francisco José Pereira Lopes UFRJ https://orcid.org/0000-0002-2715-2341
• Gilson Antônio Giraldi LNCC https://orcid.org/0000-0003-0623-9461

Resumo

This work investigates the application of deep learning techniques to analyze nanomaterial images acquired by scanning electron microscopy (SEM). The main objective is to improve the segmentation of zinc oxide (ZnO) nanoparticles and graphene oxide (GO) nanosheets. For this purpose, we developed a dedicated database for the supervised training of deep neural networks, using the U-Net architecture. Given the high spatial resolution of the original images, we investigated two approaches for image subdivision during the training and testing stages. Due to the high cost of acquiring and annotating SEM images of nanomaterials, we applied transfer learning through fine-tuning, adapting the network previously trained on ZnO images to segment GO images. The results were promising, with accuracies and recalls exceeding 95% for both ZnO and GO images.

Palavras-chave: Deep learning, U-Net, nanomaterial, scanning electron microscopy