Human vision is a powerful sensory system for detecting environmental information. However, its efficiency comes along with built-in fallacies often neglected. Utilizing a machine-learning approach, we reveal a universal hidden structure embedded in most natural images and show that 2D natural images can be compressed and thus encoded faithfully by vorticities along 1D boundaries. In addition, a hierarchy of visual information can be constructed according to the human-eye sensitivity. By projecting out the principal components for human vision, the invisible textures of the natural images emerge, providing a promising tool for medical image analysis in the future.