Shape Descriptors Based on Extended 3D Higher Order Local Autocorrelation Masks

Abstract

Mask patterns based on Higher Order Local Autocorrelation (HLAC) have been used for various 2D image recognition and classification systems. Recently, research related to extended 2D HLAC mask patterns has shown that the use of mask patterns improves classification rates for certain 2D textures. We have applied a similar extension approach to 3D HLAC mask patterns which can analyze 3D voxel data. Since there are a large number of combinations for 3D HLAC mask patterns compared with the number of 2D HLAC mask patterns, some mask patterns have to be eliminated by using proper weighting algorithms. We have used a technique based on backpropagation networks for weighting shape descriptors which are extracted by extended 3D HLAC mask patterns. The backpropagation networks were trained by learning voxel data sets, and the network eliminated unnecessary shape descriptors. A database of artificially generated 3D voxel data was used for testing pattern classification efficiencies of the extended 3D HLAC mask patterns. Our preliminary experiments showed fair recall-precision results for retrieving similar 3D voxel data when extended 3D HLAC mask patterns are used in conjunction with backpropagation networks.