The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-IPI-1040

“Structural Analysis of Natural Textures”

by Felicia Mary D'Auria Vilnrotter

September 1981

In this dissertation a general method is presented for the structural analysis of natural texture images. Edge Repetition Arrays (ERAs) are calculated for the edge and direction images corresponding to the natural texture image being analyzed. Each ERA entry is the normalized frequency of occurrence of specific types of edge matches occurring with a particular angle and distance separation. The two types of edge matches sought are for element size and spacing. Hence, a one-dimensional structural texture profile can be inferred from the information inherent in these arrays. An algorithm designed to automatically interpret ERA results is presented. This algorithm produces a one-dimensional structural description of the texture, and provides a starting point for the two-dimensional texture primitive search.

A texture primitive extraction algorithms is also developed. It uses the information inherent in the above-mentioned texture description to identify the locations of various types of elements within the texture image. The results produced are in the form of individual and composite texture primitive masks as well as descriptions for the individual texture primitive types.

A structural texture description should include information about the manner in which the primitives are arranged within the texture. Since the texture primitive masks pinpoint the texture element locations they can be used to calculate a set of rules which characterize this arrangement. An algorithm is developed which extracts the most frequently occurring interprimitive placement rules, and in case the texture being analyzed is homogeneous and regular a minimum set of rules necessary to characterize the underlying texture pattern is also chosen. Finally, a texture reconstruction algorithm is developed to produce a reconstruction of the original texture using the minimum placement rule set and a set of average primitive templates. In this way, a visual comparison can be made to determine how closely the grid relations chosen capture the spatial relations of the original texture image.

A scheme is presented for texture classification (of 11 texture types) using the structural texture descriptions discussed above. An overall success rate of 91.23% was achieved. The techniques developed in this thesis are also applied to the surface orientation determination problem. A general algorithm is outlined for determining the orientation of textures surfaces. Preliminary results are presented and discussed.

To download the report in PDF format click here: USC-IPI-1040.pdf (33.3Mb)