Capture and transparency in coarse quantized images.

Abstract

This study examines the effect of coarse quantization (blocking) on image recognition, and explores possible mechanisms. Thresholds for noise corruption showed that coarse quantization reduces drastically the recognizability of both faces and letters, well beyond the levels expected by equivalent blurring. Phase-shifting the spurious high frequencies introduced by the blocking (with an operation designed to leave both overall and local contrast unaffected, and feature localization) greatly improved recognizability of both faces and letters. For large phase shifts, the low spatial frequencies appear in transparency behind a grid structure of checks or lines. We also studied a more simple example of blocking, the checkerboard, that can be considered as a coarse quantized diagonal sinusoidal plaid. When one component of the plaid was contrast-inverted, it was seen in transparency against the checkerboard, while the other remained "captured" within the block structure. If the higher harmonics are then phase-shifted by pi, the contrast-reversed fundamental becomes captured and the other seen in transparency. Intermediate phase shifts of the higher harmonics cause intermediate effects, which we measured by adjusting the relative contrast of the fundamentals until neither orientation dominated. The contrast match varied considerably with the phase of the higher harmonics, over a range of about 1.5 log units. Simulations with the local energy model predicted qualitatively the results of the recognizability of both faces and letters, and quantitatively the apparent orientation of the modified checkerboard pattern. More generally, the model predicts the conditions under which an image will be "captured" by coarse quantization, or seen in transparency.