Wednesday, April 12, 2017
10:00 am in Rice 404
Committee: Connelly Barnes (Advisor), Vicente Ordonez-Roman (attending faculty member)
Title: Modeling Human Perception of Inpainted Images
In this project we present several experiments which investigate how humans perceive edited images. We focus on images which have been modified through the use of various image inpainting techniques. Our work follows three primary strands of inquiry: first, we compare the results of fully-automated image completion techniques with artificial images which are manually generated by humans. We then generate a set of edited images using several inpainting techniques, and analyze how well humans are able
to detect inpainted content in this dataset. Finally, we use our resulting set of realism annotations to train various deep learning systems which emulate how humans perceive composite realism. We evaluate the performance of several machine learning architectures, and show that our best models are able to perform better than baseline classifiers at the task of modeling user deductions regarding composite realism. We
conclude by analyzing potential applications, limitations, and avenues for future work which stem from this project.