In experiments, the original as well as partially processed image have superimposed noise. Hence the distribution of intensity across fringes will be ambiguous. Specifically ,the minimum intensity will not strategy be a zero and 8-bit digitization, the maximum will not be 225. Hence the strategy that is preferred is to trace the minima within dark bands and maxima within bright bands, rather than search for points of known intensity. This is the closest approximation that can be achieved in tracking a series of pixels having low-or high-intensities. When a laser is used, one must also take into account the overall Gaussian profile of the light output. Hence, determination of extrema in  intensity becomes a local operation in the image domain. Subsequently, the locus of minima or need to be connected within a fringe band across the image, to get a curve on which temperature itself or a temperature-dependent function is a constant.

It is clear that under experimental condition, only intensity minima can be traced since intensity itself is not precisely defined. In practice, one observes a greater noise level in the high-intensity regions, possibly owing to device saturation in the recording medium. This makes locating intensity maxima a difficult task. Hence, fringe thinning operations referred in the present work are related solely to the location of intensity minima in the dark fringe bands.

Several fringe thinning methods are available in the literature. Most of the fringe tracing algorithms that have been suggested are problem–specific and cannot be accepted as generally valid. A large number of published algorithms are based on edge detection by global thresholding, but these are specific to a class of problems and appear to be inapplicable for interferometric images. In the presence of an non-uniform average level of illumination with superimposed noise, the opinion of the present author is that the task of automatic extraction of the fringe skeleton is difficult. This has been experienced even with the well-behaved fringe tracing algorithms given by Robinson [83] and Krishnaswamy [84] for interferometric fringes and by Ramesh and Singh [85] for photoelastic fringe patterns. Funnell [86] has sugessted an easy to implement, but not a fully automatic, technique to trace the fringes. His method is likely to yield good results for low-quality images. In the present study, an algorithm that is similar in approach has been proposed. This algorithm that is similar in approach has been proposed. This algorithmis automatic in the sense that no user input is required at any intermediate stage of the calculation. It is based on the actual two-dimensional grey-level variations and the fringe skeleton is traced by searching alng the minimum intensity direction while simultaneously maintaining connectivity of the point traced.

In the present section, the performance of the fringe thinning algorithms has been evaluated in the context of interferometric image. These images were generated by a Mach-Zehnder interferometer. The experiment performed was one involving Rayleigh-Benard convection. The experiment comprised of a layer of air confined between two horizontal surfaces. The lower surface was heated while the upper surface was cooled, both being maintained respectively at constant temperatures. The vertical side walls that defined the boundaries of the fluid layer were thermally insulated. The temperature differential across the fluid layer led to buoyancy-driven motion, whose pattern could be captured from the fringe patterns. This experiment is described in detail in section 7.3.

The interferograms represent the projection of the three-dimensional temperature field on to a plane. However, the data contained in them must be transferred to a uniform grid before tomographic algorithms can be applied. Several intermediate steps involving image processing operations have to be performed to reach this stage. To start with, interferograms recorded using the CCD camera contain superimposed noise, speckle being the most significant. While speckle is associated with the optical components of the interferometer, the images also carry thermal noise due to edge effects in the fluid layer. Speckle, as well as thermal noise can be conveniently removed in the Fourier domain by a band–pass filter. An example of the original intensity distribution and that obtained after Fourier-filtering is shown in figure 4.16. The filter is two-dimensional and can remove the high-wave number components of the spectra of the light intensity. The resulting image has blurred edges and must be processed further. The image quality of the filtered image has been enhanced by using utilities such as histogram equalization and high-boost image preparation.