A novel methodology to identify a fatigue crack propagation law is proposed. It is based on an optimization scheme that directly interrogates the images from the sought law rather than using a sequence of fitting schemes. Two procedures are presented and tested first on artificial data with a known law, and second on an actual image series obtained from a fatigue test performed on a thin titanium sheet. The results demonstrate the ability of the proposed procedures to deal with inherently noisy data, and yet obtain an accurate propagation law. As a side result "admissible" time series of stress intensity factors and crack advance are obtained that fulfill the identified law and are constrained by the registration of actual images, rather than being simply filtered from the data in a post-processing step.