When
Where
Title
Image Classification with Convolutional Neural Networks in the Presence of Optical Aberrations and Stray Light
Abstract
As optical systems advance in capability, they present new engineering challenges. This work focuses on two such challenges: controlling stray light in imaging systems with increasingly demanding out-of-field stray light requirements and understanding the relationship between imaging hardware requirements and image classification performance with Convolutional Neural Networks (CNNs). These topics are unified by investigating CNN classification performance in the presence of stray light from optical scatter.
The first focus of this work explores the out-of-field stray light performance metric, Point Source Transmittance (PST). Sunshades are an important stray light control that improve the PST of an optical system by forcing additional scattering events to occur before stray light can reach the detector. Sunshade design is discussed to show that increasing stray light suppression requires increasing sunshade volume. While PST is a fundamental metric quantifying an optical system’s out-of-field stray light performance, imaging systems often have the concern of stray light from extended sources like the Sun, Moon, and Earth in their scene. A method of using a PST to calculate an extended source transmittance for an arbitrary source is also discussed. This approach enables flexibility in how stray light requirements are defined and managed on an effort.
Transitioning to the second focus of this work, in order to understand how CNNs are impacted by imaging hardware requirements, CNNs pretrained on pristine imagery are given imagery degraded by progressive amounts of optical aberration and correlated to the consequent degradation in classification performance. The Modulation Transfer Function (MTF) proves a critical image quality metric, as CNNs learn and recognize spatial frequency content of their targets. The layered structure of CNNs, in which successive operations reduce the spatial resolution of the input image, introduces a bias toward learning lower spatial frequency content. This is supported by classification performance with various optical aberrations; CNNs can be particularly robust against degradation with optical aberrations (that did not exist in the training imagery), so long as they do not introduce a zero value in MTF or significant degradation in contrast for the spatial frequency content of features that were learned for a given target. This discussion is extended to the stray light mechanism of optical scatter. CNNs are generally robust against optical scatter, but an important exception is when bright sources are present in a scene. Optical scatter (and other stray light mechanisms) in the presence of bright sources can raise the noise background or saturate imagery and degrade the contrast and apparent color of targets for classification.
Please email Jini at jini@optics.arizona.edu or Page at pking@arizona.edu for a Zoom link.