Evaluating Reconfigurable Hardware for Accelerating Industrial CT

Abstract

Industrial Computed Tomography (ICT) has a potential for improving processes in such areas as manufacturing, electrical and electronic devices, inhomogeneous materials, and the food industry. To be effective and scalable in industrial settings, however, its implementation must meet crucial constraints, particularly including fast response matching the short cycle times and throughput levels required, for example, by manufacturing applications. One possible bottleneck for ICT is the inherent high-performance computing demand posed by image reconstruction, an important step of scanner data processing. This paper presents the development of an FPGA-based Maximum Likelihood Expectation Maximization (MLEM) kernel, an iterative algorithm used for image reconstruction. We rely on an OpenCL-based design flow and explore a set of optimizations applied through high-level code. The results show that a carefully designed OpenCL-based accelerator can achieve performance gains as high as 8X against an unoptimized design.