A Bayesian segmentation approach to 3-D tomographic reconstruction from few radiographs

A MAP (maximum a posteriori) 3-D reconstruction technique for estimating a solid object directly from sparse cone-beam data has been developed. In the present work, emphasis is placed on radiographic flaw detection in solid materials, which can be viewed as a segmentation of the object into a binary-valued reconstruction. Optimization is performed by iterated conditional modes, with deterministic convergence, but a solution dependent on the initial condition. To speed convergence and improve the estimate, an initial condition based on maximum-likelihood estimation of flaw location is used. This MAP tomographic estimation algorithm provides a simple, robust method for segmenting a 3-D object from digitized radiographs. Its output is appropriate for either automated decision-making or visual inspection, but avoids the necessity of making decisions independently on separate radiographs, as its currently typical in application. Although the estimation process is computationally costly in terms of cost per pixel at each iteration, its convergence is very rapid in terms of iteration counts.<>