Multiresolution pyramids for segmentation of natural images based on autoregressive models: application to calf leather classification

Classification of natural surfaces for quality inspection is a step in the automation of some factory functions in industry. It is noted that, in cases of surfaces represented by random field images, some of their parameters can be useful for pattern recognition and texture analysis. Assuming the parameter values to be governed by some probability density law which takes into account the spatial distribution, a data structure can be built, starting at the original image, by gathering data properly, in a son-father order, to construct reduced-resolution versions of that image in an exponentially tapering pyramid of arrays of sizes 2/sup n/*2/sup n/, . . ., 2*2. Iterative pyramids are then built to update the node values based on their sons and fathers. Nodes are enhanced on different levels that feature no relationship to their parents and become probable roots for regions to be segmented on the lowest levels. A pyramid linking process then follows the root identification. This methodology was implemented and applied to segmenting some defects in calf leather.<
>