Multistep Detection of Oriented Structure in Complex Textures
Abstract
In the context of online industrial control, we propose a multistep method, based on mathematical morphology operators, to carry out the characterization of oriented structures in a complex texture environment. In a data base of about ten thousand patches, ten percent show various oriented structures. First we measure for each patch the x and y covariances, and apply pyramids of dilations-erosions and openings-closings using linear structuring elements. Second we transform the resulting pyramid images into usable condensed information. All the measures for each patch set up texture descriptors. For each of them we apply a Principal Components Analysis (PCA) to sum up the information. Then we use a Linear Discriminant Analysis (LDA) on the principal axes which carry the most information and analyze in the LDA space the distribution of the two types of patches to reach 99.45% of sensitivity and 99.61% of specificity.
Origin : Files produced by the author(s)
Loading...