Filtering of feature matches is heuristic method aimed to reduce the number of feasible matches and is widely employed in different image registration algorithms based on local features. In this paper we propose to interpret the filtering process as an optimal classification of the matches into the correct or incorrect match classes. The statistics, according to which the filtering is performed, uses differences of the geometrical invariants obtained from ordered sets of local features (composite features) of proper cardinality. Further, we examine some computationally efficient implementation schemes of the classification. Under the assumption of Gaussian measurement error, the conditional distribution densities of invariants can be approximated by well-known linearization approach. Experimental evidences obtained from fingerprint identification, which confirm viability of the proposed approach, are presented.