Raw data from a digital imaging sensor are impaired by a heteroscedastic noise, the variance of pixel intensity linearly depending on the expected value. The most natural way of estimating the variance and the expected value at a given pixel is certainly empirical estimation from the variations along a stack of images of any static scene acquired at different times under the same camera setting. However, the relation found between the sample variance and the sample expectation is actually not linear, especially in the presence of a flickering illumination. The contribution of this paper is twofold. First, a theoretical model of this phenomenon shows that the linear relation changes into a quadratic one. Second, an algorithm is designed, which not only gives the parameters of the expected linear relation, but also the whole set of parameters governing an image formation, namely the gain, the offset and the readout noise. The rolling shutter effect is also considered.