Human and DNN Classification Performance on Images With Quality Distortions

Abstract

Image quality is an important practical challenge that is often overlooked in the design of machine vision systems. Commonly, machine vision systems are trained and tested on high-quality image datasets, yet in practical applications the input images cannot be assumed to be of high quality. Modern deep neural networks (DNNs) have been shown to perform poorly on images affected by blur or noise distortions. In this work, we investigate whether human subjects also perform poorly on distorted stimuli and provide a direct comparison with the performance of DNNs. Specifically, we study the effect of Gaussian blur and additive Gaussian noise on human and DNN classification performance. We perform two experiments: one crowd-sourced experiment with unlimited stimulus display time, and one lab experiment with 100ms display time. In both cases, we found that humans outperform neural networks on distorted stimuli, even when the networks are retrained with distorted data.