Second-harmonic generation (SHG) can be used for imaging membrane potential in neurons, but poor signal-to-noise (S/N) limits accurate measurements of small voltage transients. We use photon counting to improve the S/N of weak SHG signal detection. Photon counting generates shot-noise limited and integrable signals, eliminates pulse-to-pulse variation, and built-in discriminators reduces the background to practically zero. In single trials, by using photon counting, we obtain a more than a twofold S/N increase over analog voltage detection. Trial-to-trial variability is also reduced by 50%. Finally, we show that, using photon counting, the kinetics of fast events such as action potentials can be recorded more accurately.