This paper discusses a two-level approach to optimize classification systems with multi-objective genetic algorithms. The first level creates a set of representations through feature extraction, which is used to train a classifier set. At this point, the most performing classifier can be selected for a single classifier system, or an ensemble of classifiers can be optimized for improved accuracy. Two zoning strategies for feature extraction are discussed and compared using global validation to select optimized solutions. Experiments conducted with isolated handwritten digits and uppercase letters demonstrate the effectiveness of this approach, which encourages further research in this direction.