One way to extend depth of field of any optical system is to take several images with different focus positions and combine them into a single composite image, which contains all regions fully focused. The challenge then becomes to select from each image the pixels that are in focus. We describe a new focus measure based on the shapelet decomposition. Results using simulated images with high noise content show that the shapelet-based focus measure is the most performance that traditional neighborhood-based focus approaches. Similarly shapelet focus measure provides significant improvement compared to traditional methods when objects have nontextured or homogeneous regions.