Recently we have leveraged the statistics of natural images to learn a low-dimensional computational model of visual texture selectivity (Oleskiw & Simoncelli, SfN 2018). Our model, trained to detect naturalistic statistics across stimuli matched in local orientation energy, learns computational units resembling localized differences (i.e., derivatives) spanning the 4-dimensional space of V1 selectivity, namely horizontal and vertical position, orientation, and scale.

In the present work, we use this observation to construct a population of V2-like units that captures statistics of natural images. Images are first encoded in the rectified responses of a population of V1-like units localized in position, orientation, and scale. We then explicitly compute derivatives of local spectral energy via linear combinations over the 4D space of V1 activity. A low-dimensional population of simple (rectified linear) and complex (squared) V2-like units is shown to accurately distinguish natural textures from spectrally matched noise, outperforming even the optimal Fisher linear discriminant trained over the full set of V1 afferents (91% vs. 86% accuracy, using only 0.2% of available dimensions). We conclude by demonstrating how this canonical and physiologically-plausible model of V2 computation selectively captures complex features of natural images from the local spectral energy conveyed by V1.