SPECTRAL AND RANK ORDER APPROACHES TO TEXTURE ANALYSIS

There are two major approaches to texture analysis, both supported by physiological evidence: those based on the spatial statistics, and tho se based on its spectral properties. One of the most sophisticated spe ctral approaches to texture is that based on the Wigner distribution w here the attributes computed for each pixel encapsulate both the local spectral and phase properties of the local Fourier transform in a uni que real spectrum. On the other hand, some of the most efficient metho ds which operate in the spatial domain alone, are those based on rank order functions. Before one embarks on the use of the sophisticated me thods, it is worth exploring the efficient ones to the limit of their performance. In this paper we investigate these two major approaches a nd compare their performance both in terms of quality of results and e fficiency. The problem we consider is that of detecting defective blob s and cracks on complex textural backgrounds. We show that in most cas es rank order approaches can perform well, although no unique method c an be employed for both types of defects. On the other hand, the Wigne r approach with a very small modification can cope with both types of defects and handle even the identification of very subtle cracks. Thus , it seems that for any real time performance inspection system, the r ank order approaches should form the front end with the more sophistic ated methods coming in play when necessary.