Reconstruction with 4 view angles is taken up next. Table 15 shows the error levels in the reconstructed data and the distribution of this error within the fluid layer. It can be seen immediately that the errors shows that these are at best localized, i..e., large errors may occur at a few sporadic points. The AVMART1 algorithm shows a considerable deterioration in performance, since errors as well as CPU time are substantially higher. AVMART2 and AVMRT3 algorithms are seen to perform

Table 15: Comparison of the AVMART Algorithms: Noise in projection data, Four-View Reconstruction  

Quantity	AVMART1	AVMART2	AVMART3
	11.80	5.52	5.52
	1.78	1.36	1.36
	10.41	8.00	8.00
Number of Points (%) having error int he range
>95	0.004	0.007	0.007
75-95	0.029	0.349	0.346
50-75	0.276	5.186	5.177
Inerations	190	53	53
CPU,sec	1767.7	502.3	520.8

Better than AVMART1. AVMART2 is marginally superior to.  AVMART3 since the error magnitudes are equal, but the former takes a smaller CPU time. Hence, a consolidated view to emerge from the discussion above is that AVMART2 exhibits the best performance.
It is of interest to compare the best proposed algorithm, namely AVMART2 with the best original MART algorithm identified by Subbarao et al. [32], Namely MART3 of the present study. To this end, reconstruction was carried out using 2-views of for convection in a horizontal differentially heated fluid layer, leading to two- dimensional longitudinal rolls. The projection data was superimposed with noise and an initial guess of a constant temperature field was used. Errors for MART2–new. The computer time was also higher by a factor of4 when compared with MART2_new. However the fractional distributions of errors over the fluid layer were seen to be similar for both, thus confirming that they continued to belong to the same family of algorithms.
The following inferences can now be drawn from the discussion above:
1. The three AVMART algorithms show similar performance in the presence of noise in the projection data AVMART2 is however marginally superior in terms of errors and CPU time.
2. The noise in the projection data persists after reconstruction.
3. Increasing the number of noisy projections amplifies the error in reconstruction.
4.  AVMART2 clearly shows superiority over MART3 for noisy projection data. Hence it supersedes MART3 as the favored tomographic algorithm for the class of problems studied.
Both MART and AVMART algorithms have been tested extensively against experimental data. The errors as well as the convergence rates have been tested extensively against experimental data. The errors as well as the convergence rates have been reported in
Mishra et al. [81]. The conclusions drawn above carry over to experiments without any major modification. The convergence rates of all the algorithms were seen to deteriorate with increasing number of projection angles. This could be traced to the partial de-correlation among the interferometric  images owing to mild unsteadiness in the convection patterns.