Frequency-domain deconvolution in deep learning

This dissertation presents an exhaustive exploration of a novel approach to deep learning in computer vision tasks: the frequency-domain deconvolution operation. Recognizing the unparalleled success of convolutional neural networks (CNNs) in the realm of computer vision, we critically evaluate the performance and computational demands of traditional convolution operations against the proposed deconvolution method. Using a systematic approach, we apply the deconvolution layer to two quintessential computer vision problems: image classification and single image super resolution (SISR). The results demonstrate the deconvolution layer's potential in certain scenarios, with marked improvements observed in image classification. For SISR tasks, though advantages were noticed under specific configurations, the traditional CNNs still demonstrated their robustness. Additionally, the dissertation touches upon the layer's computational demands, revealing an increased computational overhead for the deconvolution layer. Encouragingly, the layer demonstrated promising attributes like learning long-range filters and isolating objects from backgrounds effectively. Concluding with avenues for future research, this dissertation acts as a stepping stone in the uncharted territory of deconvolution operations, emphasising innovation alongside evaluation in the dynamic world of deep learning.

Grantor dc:publisher.institution

Department of Electrical Engineering

Year dc:date.issued

2025

Author dc:creator

• Meyer, Moegamad

Advisor dc:contributor.advisor

• Nicolls, Frederick

• Convolutional neural networks

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/11427/41745

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/41745