In this paper, we present a new method to select the faint, background galaxies used to derive the mass of galaxy clusters by weak lensing. The method is based on the simultaneous analysis of the shear signal, that should be consistent with zero for the foreground, unlensed galaxies, and of the colours of the galaxies: photometric data from the COSMic evOlution Survey are used to train the colour selection. In order to validate this methodology, we test it against a set of state-of-the-art image simulations of mock galaxy clusters in different redshift [0.23-0.45] and mass [0.5-1.55 x 10(15) M-circle dot] ranges, mimicking medium-deep multicolour imaging observations [e.g. Subaru, Large Binocular Telescope]. The performance of our method in terms of contamination by unlensed sources is comparable to a selection based on photometric redshifts, which however requires a good spectral coverage and is thus much more observationally demanding. The application of our method to simulations gives an average ratio between estimated and true masses of similar to 0.98 +/- 0.09. As a further test, we finally apply our method to real data, and compare our results with other weak-lensing mass estimates in the literature: for this purpose, we choose the cluster Abell 2219 (z = 0.228), for which multiband (BVRi) data are publicly available.