Many image processing applications require to detect a known pattern buried under noise. While maximum correlation can be implemented efficiently using fast Fourier transforms, detection criteria that are robust to the presence of outliers are typically slower by several orders of magnitude. We derive the general expression of a robust detection criterion based on the theory of locally optimal detectors. The expression of the criterion is attractive because it offers a fast implementation based on correlations. Application of this criterion to Cauchy likelihood gives good detection performance in the presence of outliers, as shown in our numerical experiments. Special attention is given to proper normalization of the criterion in order to account for truncation at the image borders and noise with a non-stationary dispersion.