Locally Adaptive Techniques Forstack

This paper introduces a new structure for stack lter-ing, where the lter adapts to the local characteristics encountered in data. Both supervised and unsupervised techniques for optimal design are investigated. We split the image into small regions and select the stack lter to process each region according to the spatial domain or threshold level domain characteristics of the input signal. This method provides a signiicant improvement potential over the global stack ltering approach. Some local statistics are computed, to build a reduced input space which eeciently describes the most important local characteristics of data. Vector quantization is used for clustering the reduced input space into a small number of regions, and then nding a mapping between reduced input space clusters and the lter space, will result in rules for selecting the best suited stack lter for a given region. The supervised clustering procedures are shown to surpass signiicantly the global ltering approach.