Graphics Gems II, chapter II.9 (Image File Compression Made Easy) says:
These constraints are easily met for arithmetic prediction using Roberts’s method L.R., a “poor man’s Lapacian” (gradient estimator), which finds the slope along the NW/SE diagonal: G(a, b, c) = b + c – 2a. Adding this slope step to the variable a yields the estimate: (*) = a + G(a, b, c). The predictor thus is P(a, b, c) = b + c – a. For images that ramp in linear fashion along any axis, the prediction is exact. Viewing the predictor output for an entire forms an image reminiscent of an unsharp mask convolution (Hall et al., 1971).
But I can't figure out where the mentioned "Roberts's method" for gradient estimation came from: it turns out there are several mathematicians named Roberts, which makes tracking down the one mentioned hard. Where did Roberts originally introduce his method for gradient estimation?
It was originally described in Machine Perception Of Three-Dimensional Solids by Lawrence G. Roberts as a method for creating differential images to do edge dection, and is also called the Roberts cross differential operator.