The Image flip effect on a CNN model classification

Abstract

Convolutional neural networks (CNNs) have been successful applied to many fields such as image processing. The CNNs can automatically extract image features and make corresponding predictions. However, some minor changes on the images can lead the neural network to wrong predictions. In this paper, we will use image flip technique to evaluate a CNN performance. The experiments show the model are sensitive to a particular flip direction. Additional training with the preprocessing technique have been proved helpful in improving the robustness of the model.