Differential testing · Image processing · CUDA
Auditing the imaging front end, stage by stage
A differential audit of a scanner's image-processing front end against two independent reference implementations — closing numeric parity and surfacing discrepancies that aggregate quality metrics had averaged out of sight.
Challenge
Three implementations of the same front end disagreed in ways that had never been quantified. Small systematic biases in centroid extraction become depth errors downstream, and aggregate quality metrics average them out of sight.
Contribution
Compared every stage numerically across implementations — filter outputs, sub-pixel centroids (feature positions located to a fraction of a pixel), decoded pattern codes, depth. Isolated a +1.0 px systematic centroid bias, roughly 220 µm of depth error, in a recursive smoothing filter: a Deriche filter runs each image line forward and then backward, and the backward leg was off by one sample. Also found a scan path that never re-merged pattern data split across alternating frames.
Outcome
Numeric parity closed to 98.9% of 1.1 million centroids within 1 µm, and re-merging the split pattern path cut core noise by 13% while raising valid points per frame by 74%.
Client work is described at a technical, client-agnostic level.