Date of Award
Campus Access Thesis
Master of Science (MS)
Dan A. Simovici
Humans perceive their surroundings in great detail even though most of our visual field is reduced to low-fidelity input by the retina. In contrast, most deep learning architectures are computationally wasteful in that they treat every part of the input image as equally important to the underlying image processing task. We know that the human visual system is able to perform visual reasoning despite having only a small fovea of high visual acuity. Attention is the mechanism by which our visual input is filtered and our visual system's limited computational resources are allocated. Interest in incorporating attention mechanisms into existing deep architectures has recently been renewed. In support of such research, we investigate the extent to which existing architectures are able to learn useful feature representations from low acuity or fidelity inputs. Specifically, we apply deep autoencoders to the problem of image super-resolution and find that they can learn remarkably rich hidden representations from low-resolution inputs such as color and shape features from natural images and strokes from digit images. We find that in many cases, the learnt features are suitable for reconstructing the original high resolution image.
Ahmed Wick, Farahnaz, "Filling in the Details: Perceiving from Low Fidelity Images" (2016). Graduate Masters Theses. 404.