We present a speckle order estimation method invariant to mean intensity variations inside the sample. It is based on the acquisition of two statistically independent speckle images. The Cramer-Rao lower bound and the maximum likelihood estimator in the presence of pure speckle noise are derived and shown to have performance very close to classical estimation methods that are not invariant to mean intensity fluctuations. This method is also shown to be robust to Poisson shot noise and additive noise, and is validated on an optical experiment with a standard camera. Thus, it is useful for estimating speckle order in real-world coherent images.