Date of Graduation

8-2024

Document Type

Thesis

Degree Name

Master of Science in Statistics and Analytics (MS)

Degree Level

Graduate

Department

Statistics and Analytics

Advisor/Mentor

Kaman, Tulin

Committee Member

Zhang, Qingyang

Second Committee Member

Plummer, Sean

Keywords

GAN model; Discriminator architecture; Skip patch

Abstract

GAN models have been successfully used for image generation in various sections such as real-life objects like human faces, cars, animal faces, landscapes, etc. This work focuses on biological electron microscopy (EM) image generation. Unlike other real-life objects, biological EM images are obtained through electron microscopy techniques to study biological specimens. Electron microscopy offers high resolution and magnification capabilities, making it a powerful tool for visualizing biological structures at the nanoscale. However, using GAN models for biological EM image generation poses challenges due to the complex and unique arrangements of biological structures and the sparse and asymmetrical patterns in EM images, making it difficult for the model to generate realistic images accurately. The patch-based GAN discriminator lacks the capability of simultaneously accessing both the global and local structures of the generated image. The patch discriminator operates at a small receptive field (16x16 patch) in capturing precise local structures while struggling to represent global structures accurately. Conversely, the patch discriminator equipped with large receptive fields (70x70 patch) effectively captures global structures but often fails to reproduce detailed local textures accurately. By addressing the challenges, I have proposed a new discriminator architecture for training GAN models in settings with limited data availability and the presence of both global and local structures to generate realistic biological EM images. The proposed architecture is called a skip patch discriminator.

Download

Share

COinS