FPGA-accelerated Automatic Alignment for Three-dimensional Tomography

In the process of tomographic reconstruction, the attitude and center point of a specimen, from which the projection data are collected, suffer from misalignment due to mechanical imperfection and calibration error. Such misalignment leads to poor reconstruction quality. And effective automatic alignment approaches have been proposed. The alignment approaches are of good use for kinds of application scenarios such as X-CT and electron tomography. These scenarios demand not only high performance, but also that the component of automatic alignment can be integrated and upgraded in the whole solution. Thus, we propose an FPGA accelerator for state-of-the-art tomographic alignment algorithm. We first introduce a multi-ray access approach that modifies the order of data access for easier on-chip data management. Making use of BRAMs on FPGAs and effective local data management strategy, data reuse is reinforced, and data transfer latency with DRAM is covered by computation. Also, we introduce an FPGA-customized processing engine at a low cost to improve data throughput. Moreover, a streaming structure with multiple paralleled PEs further improves the performance of our algorithm. Experiments demonstrate that our accelerator achieves a 44. 5x speed-up for the state-of-the-art alignment on Xilinx ZCU102 over a 16-thread multicore CPU implementation, and a 1. 60x speed-up with 7. 8x energy reduction over an OpenCL implementation on Nvidia Titan V.