Optimization of the constrained Algebraic Reconstruction Technique for the performance of a variety of visual tasks

This work demonstrates the application of a method to optimize image reconstruction algorithms on the basis of the performance of specific visual tasks that are to be accomplished using the reconstructed images. The evaluation of task performance is numerically realized by a Monte Carlo simulation of the complete imaging chain, including the final inference based on the reconstructions. Fundamental to this evaluation is that it yields an average response by consideration of many initial scenes. It is shown that the use of the nonnegativity constraint in the Algebraic Reconstruction Technique can significantly improve performance in situations where there is a severe lack of measurements when the relaxation factor is optimized. There is no indication in any of the cases studied hitherto that the nonnegativity constraint can improve performance in situations where the data are complete, but noisy. It is also shown that neither the rms error nor the L1 error in constrained reconstructions are good predictors of task performance.