One of the major problems faced by deep learning techniques--for example, deep convolutional neural networks (CNNs)--is uncontrolled behavior due to glitches in the input. For example, you can modify just a few pixels in a photo’s background and the neural network won’t be able to recognize it anymore. CNNs are less affected by changes to an image via color or contrast; however, an increase in blur or noise level negatively affects performance. Scrambling or cutting block pixels out from an image also worsens CNN performance, especially for indoor scenes. In general, humans are very tolerant of such image modifications.

The authors of this paper study this problem and help readers to understand it better. To do so, they compare the classification results of distorted images made by human subjects--real indoor or outdoor scenes--to the classifications made by CNNs. Furthermore, they compare the network’s classification accuracy to the accuracy of other classic image classifiers based on hue saturation and GIST descriptors. Thus, CNN is compared to five classifiers: human subjects, PlacesNet, InOutNet, GIST-LDA, and HSV-LDA. Different classification agreements between some and all of them are induced.

This paper targets a hot topic in image recognition--CNNs are being considered in biological vision research and in real-world scene analysis (like autonomous robots). This is a very-well-written scientific paper that is easy to read due to its clear and consistent explanatory style. It is also well organized. I highly recommend it to researchers interested in computer vision and image perception.