Industrial Computed Tomography (CT) scans increasingly demand for fast throughput time. It can be achieved by either reducing the number of projections or the X-ray dose. Lowering the X-ray dose leads to noise degradation, resulting in insufficient reconstruction quality with classical Filtered Backprojection (FBP) based algorithms. One alternative is statistical 3D post-denoising of reconstructions achieved with classical FBP-based algorithms. Total Variation (TV) regularization is a powerful denoising technique that preserves edges and details in images. To obtain desired speed and portability, OpenCL is suitable for implementing TV algorithms on GPUs.

In this work, we provide optimized 3D TV algorithms and a statistical 3D TV algorithm on multi-GPU systems for industrial CT data in OpenCL code. The numerical and GPU implementations of these algorithms are introduced and analyzed. The results show that the statistical denoising algorithm can improve the denoising performance both visually and quantitatively.