A HYPOTHESIS FOR THE COGNITIVE DIFFICULTY OF IMAGES

Xu Cheng, Xin Wang, Haotian Xue, Zhengyang Liang, Xin Jin, Quanshi Zhang

Keywords: cognitive difficulty, multi-variate interaction

Abstract: This paper proposes a hypothesis to analyze the underlying reason for the cognitive difficulty of an image from two perspectives, i.e. a cognitive image usually makes a DNN strongly activated by cognitive concepts; discarding massive non-cognitive concepts may also help the DNN focus on cognitive concepts. Based on this hypothesis, we use multi-variate interactions to represent cognitive concepts and non-cognitive concepts contained in an image, and further design a set of image revision operations to decrease the cognitive difficulty of the image. In experiments, we found that the revised image was usually more cognitive than the original one. Besides, we also discovered that strengthening cognitive concepts and weakening non-cognitive concepts could improve the aesthetic level of an image.

One-sentence Summary: This paper proposes a hypothesis to analyze the underlying reason for the cognitive difficulty of an image.