Date of Award
Master of Science
It has been shown that image segmentation models can be improved with an adversarial loss. Additionally, previous analysis of adversarial examples in image classification has shown that image datasets contain features that are not easily recognized by humans. This work investigates the effect of using a second adversarial loss to further improve image segmentation. The proposed model uses two generative adversarial networks stacked together, where the first generator takes an image as input and generates a segmentation map. The second generator then takes this predicted segmentation map as input and predicts the errors relative to the ground truth segmentation map. If these errors contained additional features that are not easily recognized by humans, they could possibly be learned by a discriminator. The proposed model did not consistently show significant improvement over a single generative adversarial model, casting doubt about the existence of such features.
Matthew John Burke
Burke, Matthew John, "Stacked generative adversarial networks for learning additional features of image segmentation maps" (2020). Graduate Theses and Dissertations. 17983.