Image representation by blob and its application in CT reconstruction from few projections

TL;DR

This paper introduces a novel image representation using blobs for CT reconstruction from limited projections, employing shift-invariant spaces and optimization techniques, with GPU acceleration for efficient computation.

Contribution

It develops blob-based image models and reconstruction algorithms specifically designed for few projection CT data, enhancing reconstruction quality and efficiency.

Findings

Effective reconstruction with few projections demonstrated.

GPU acceleration significantly speeds up computations.

Blob representation improves image quality over traditional pixel basis.

Abstract

The localized radial symmetric function, or blob, is an ideal alternative to the pixel basis for X-ray computed tomography (CT) image reconstruction. In this paper we develop image representation models using blob, and propose reconstruction methods for few projections data. The image is represented in a shift invariant space generated by a Gaussian blob or a multiscale blob system of different frequency selectivity, and the reconstruction is done through minimizing the Total Variation or the 1 norm of blob coefficients. Some 2D numerical results are presented, where we use GPU platform for accelerating the X-ray projection and back-projection, the interpolation and the gradient computations.