A compact perceptual space for natural textures emerges from natural image statistics

Natural scenes are full of textures, enabling us to recognize surface materials and bringing richness and realism to our perceptual experience. While models have represented texture perception using high-dimensional statistical features, final perceptual appearance of a natural texture may be determined by more compact neural representations. Previous attempts derived such a compact space from verbal descriptors, but it captures a semantic rather than a perceptual space. Here, we derived the space directly from natural images using an unsupervised generative model. We show that the resulting 'texture' space, with only 12-16 dimensions, was able to generate perceptual metamers, and that distances in the space accurately aligned with human perceptual similarity judgments. While individual dimensions of the space were difficult to verbally describe, specific coordinate regions in the space corresponded to semantic descriptions such as glossiness. Furthermore, VEP analysis confirmed that visual cortical responses share a similar underlying structure, which allowed us to reconstruct the original texture stimuli. These results demonstrate that rich natural texture impressions-often resistant to linguistic description-are supported by a shared, low-dimensional structure that governs perceptual similarity, semantic interpretation, and neural encoding.