Image processing and understanding human vision have shared interests and methods since the beginning of computer vision research. In recent years, deep learning and in particular convolutional neural networks (CNNs) have been applied to both, as early human vision is believed to apply filters similarly to CNNs.

CNNs trained on large image datasets have an accuracy that rivals human ability to classify images. However, human observers are tolerant to image degradation while CNN results are debatable. The authors develop an experimental protocol to test CNNs on 200 images of indoor and outdoor scenes degraded with Gaussian blur, white noise, scrambling, and occluding grids at different levels. The CNN results are compared to results of a linear classifier using only HSV (hue, saturation, value) features and a linear model built with holistic spatial features (GIST). The results confirm that CNNs perform comparably to human observers when it comes to the original images, whereas GIST and HSV descriptors produce less accurate classifiers. For degraded images, however, CNN accuracy drops much faster than human accuracy.

Conclusions about whether CNNs model human vision are derived for high degradation levels; it seems that global image features are not well captured by CNNs. Moreover, comparing the levels of agreement suggests that CNN training creates different features from those used by humans when observing degraded images.

This paper is easy to follow, applies a rigorous method, shows many interesting statistics, and gives indications for future research. It is of interest to computer and cognitive scientists.