Directional Inhibition Network (DI-Net): An Inspectable Retina-Inspired Code for Controlled GT-Isolated One-Pixel Eight-Way Direction Classification

A long-standing account of retinal direction selectivity states that asymmetric inhibition suppresses responses to motion in the null direction. We use this idea as a computational prior in the Directional Inhibition Network (DI-Net), a two-stage model for direction classification from a before/after image pair. The first stage is a deterministic, parameter-free, anti-coincidence encoder that produces eight spatial maps of local directional evidence. The second is a compact convolutional network that combines this evidence into a global direction estimate. Because only the second stage is learned, the intermediate code remains directly inspectable. Experiments use object-conditioned pairs derived from DIS5K, with controlled translations and corruption applied only at test time. For one-pixel motion, DI-Net achieves 0.993 Accuracy on clean pairs and 0.747 under 10% corruption, close to Lucas–Kanade in the same evaluation. Fixed-channel voting reduces clean Accuracy to 0.713, whereas a convolutional network trained directly on the image pair degrades much more sharply under corruption. A multi-step version of the encoder also improves direction classification for displacements from 1 to 16 pixels compared with a parameter-matched one-step control. Tests on selected DAVIS-derived pairs show no statistically resolved difference between DI-Net and the evaluated optical-flow baselines. Taken together, these results support DI-Net as an interpretable, retina-inspired computational model for controlled motion-direction tasks rather than as a physiological account of retinal processing.