Abstract

Median filters are attractive for image processing because they remove noise while preserving edges and locally monotone regions. We describe an implementation of the nonlinear ranking operation associated with these filters that takes advantage of the speed and parallelism inherent to linear optical systems. The architecture is based on decomposition of the L level quantized input image into its L component threshold images. The / th threshold image is 1 where the input image exceeds the / th quantum level and 0 elsewhere. Median filtering of the binary threshold images may be realized with a simple moving average followed by a threshold step. Superposition of the L median filtered threshold images yields the median filtered output image.

© 1985 Optical Society of America