We develop and test a method for the estimation of carbon abundance ratios, [C/Fe], and metallicities, [Fe/H], in metal-deficient stars, based on application of artificial neural networks and synthesis models to medium-resolution (1–2 Å) spectra in the region of the CH G-band feature at $\lambda$ 4300 Å. We calibrate this method by comparison to modern carbon abundance determinations for N $\simeq 125$ stars reported in the recent literature. As a first application, we estimate the abundance of carbon and [Fe/H] for the sample of 56 stars identified as carbon-rich, relative to stars of similar metal abundance, in the sample of “strong G-band” stars discussed by Beers, Preston, & Shectman.