Understanding Which Images Degrade Deep Learning Classification Performance
Seong Tae Kim, Hak Gu Kim · TECHART Journal of Arts and Imaging Science · 2021
In recent years, deep neural networks (DNNs) have shown high success and accuracy in various classification tasks such as speech recognition, image classification, segmentation, etc. However, DNNs provide low classification performance when the train and test distributions differ (i.e., out-of-distribution). In this study, we investigate each visual characteristic of the correctly classified and misclassified images (referred to as natural adversarial examples) on CIFAR10 to understand which images substantially degrade machine learning and deep learning models’ classification performances. To verify this, we divide the classification results into two groups per class: correct prediction and incorrect prediction. Then, we compare the visual appearance (e.g., image point of view) of examples of correct predictions with those of incorrect predictions. Our experimental results demonstrated that correctly classified examples have several common visual characteristics, which differ from those of natural adversarial examples.