Matched filtering is often too sensitive to within-class variations (rotation, translation, magnification, etc.) and too insensitive to between-class variations. Improvements can be sought by changing the input, changing the filter, or changing the way the data is handled. We review both systematic and ad hoc methods to accomplish those improvements. The resulting improvements make coherent optical pattern recognition far more versatile in 1980 than it was when it began in 1964.