Texture Segmentation by Evidence Gathering

A new approach to texture segmentation is presented which uses Local Binary Pattern data to provide evidence from which pixels can be classified into texture classes. The proposed algorithm, which we contend to be the first use of evidence gathering in the field of texture classification, uses Generalised Hough Transform style R-tables as unique descriptors for each texture class and an accumulator is used to store votes for each texture class. Tests on the Brodatz database and Berkeley Segmentation Dataset have shown that our algorithm provides excellent results; an average of 86.9% was achieved over 50 tests on 27 Brodatz textures compared with 80.3% achieved by segmentation by histogram comparison centred on each pixel. In addition, our results provide noticeably smoother texture boundaries and reduced noise within texture regions. The concept is also a “higher order” texture descriptor, whereby the arrangement of texture elements is used for classification as well as the frequency of occurrence that is featured in standard texture operators. This results in a unique descriptor for each texture class based on the structure of texture elements within the image, which leads to a homogeneous segmentation, in boundary and area, of texture by this new technique.