Visual illusion cognition dataset construction and recognition performance by deep neural networks

The illusion cognition of human vision is commonly regarded as a typical recognition pattern, which serves a significant role in further analysis. In this study, an illusion cognition experiment based on DNNs is designed. Wherein the dataset is constructed by modeling a series of visual illusion scenes, including 3D stereo chessboard and 2D plane contrast optical illusion scenes. The high semantic segmentation evaluation accuracy result (over 0.97) on the constructed dataset demonstrates that visual illusion scenes can be effectively recognized by current DNN models, which also reflects that the illusion cognition of human vision is not a true illusion phenomenon and should imply an unknown expressible computing logic.