The images introduced to an optical pattern recognition system are normally an amplitude representation of the original intensity images. We define a method of encoding intensity images as phase-only images, and we demonstrate that the performance of the correlation is generally substantially better than that of the correlation performed with amplitude inputs. Energy in the correlation peak is about a factor of 20 greater for the phase-encoded input image method. These performance and energy advantages are particularly useful when designing filters that incorporate various degrees and types of invariance. We give examples and numerical results for the correlation process with ternary and 15-phase-state plus zero-amplitude filters with varying spans of invariance. We suggest practical implementations for this technique.