Outliers due to occlusions and contrast and offset signal deviations notably hinder recognition and retrieval of facial images. We propose a new maximum likelihood matching score with "soft masking" of outliers which is robust in these conditions. Differences between two images are modelled by unknown contrast and offset deviations from an unknown template and by independent pixel-wise errors. The error distribution is a mixture of a zero-centred Gaussian noise with an unknown variance and uniformly distributed outliers. The matching score combines the maximum likelihood estimates of model parameters and the soft masks being produced by a simple iterative expectation-maximisation algorithm. Experiments with facial images from the MIT face database show the robustness of this technique in the presence of large occlusions.