Skip to main content
eScholarship
Open Access Publications from the University of California

UC San Diego

UC San Diego Electronic Theses and Dissertations bannerUC San Diego

Investigating the effectiveness of Log-Polar projections in conjunction with Convolution Networks

Abstract

This study explores biologically inspired transformations and training techniques for image networks, such as log-polar projections and curriculum learning, in conjunction with convolutional neural networks (CNNs). Specifically, it presents a comparative analysis of log-polar CNNs and traditional CNN architectures. The key difference in log-polar CNNs is the conversion of input images from Cartesian to polar coordinates, followed by a logarithmic transformation of the radial coordinate 'r'. Preliminary experiments indicate that log-polar CNNs exhibit enhanced robustness to rotation and scale changes at inference time when trained with log-polar transformed images. Additionally, our results highlight improved resilience to geometric distortions and specific noise types, suggesting potential for broader applications in adversarial and biologically inspired modeling tasks. This research also investigates ways to align these log-polar networks and their representations with the human visual system's learning mechanisms, aiming to achieve superior overall performance under standard conditions.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View