Deep Style Transfer for Generation of Photo-realistic Synthetic Images of CNT Forests
Prashanth Kotha, Minasadat Attari, Matthew R. Maschmann, Filiz Bunyak · 2023
Carbon nanotubes (CNTs) are promising nano-materials with diverse applications in various fields, ranging from electronics and energy storage to biomedical applications. Characterization of CNT forest structures and prediction of material properties through image analytics are critical for new material design and discovery. Artificial intelligence and machine learning (AI/ML) driven approaches offer promising solutions towards these goals. These data-driven approaches rely on large amounts of annotated data for robust training. However, growing, imaging, and annotation of CNT forests are complex, time consuming, expensive processes limiting availability of such data. In this paper, we propose a novel neural style transfer pipeline to generate photo-realistic and structure-aware synthetic images of CNT forests by imposing scanning electron microscopy (SEM) like appearance to physics-based simulation results. The proposed pipeline relies on a novel wavelet-based style transfer network and a selective decision fusion mechanism that allows better preservation of the thin curvilinear structures of CNTs, particularly their intricate details and connectivity, without introducing unwanted artifacts. Experimental results and quantitative analysis demonstrate refined style transfer that generates photo-realistic images that can be used to train AI/ML- driven CNT forest image analysis systems.