Session: 20-01: Online NDE techniques for smart manufacturing

Paper Number: 98387

98387 - Evaluation and Comparison of Two Deep-Learning Strategies for On-Line X-Ray Computed Tomography 

X-ray Computed tomography (CT) is increasingly used in many industrial domains for its unique capability of controlling both the integrity and dimensional conformity of parts. Still, it fails to be adopted as a standard technique for on-line monitoring due to its excessive cost in terms of acquisition time. The reduction of the number of projections, leading to the so-called sparse-view CT strategy, is therefore one of the main challenges in this field. The fast and classical approach for CT reconstruction, filtered back-projection (FBP), shows limitations in sparse-view configurations and results in a degradation of the reconstructed image due to streaking or blurring artefacts. Iterative approaches on the other hand are able to reconstruct good quality images but at the cost of high computation. More recently, guided by the effectiveness of deep learning in computer vision tasks, many data-driven approaches have also been proposed for the reconstruction of computed tomography images.

Most of the applications in the literature are based on medical images and few results have been presented on NDE datasets, whose specific features in terms of structure and contrast are quite different. In this paper, we compare two different approaches for the reconstruction from few views, based respectively on pre-processing and post-processing strategies. The pre-processing strategy takes over classical interpolation schemes used in CT reconstruction, which present limitations in extremely sparse configurations. Starting from a sparse sinogram data, the objective is to learn how to complete the missing views and obtain a dense sinogram on which is performed fast and accurate reconstruction. The post-processing strategy relies on an opposite scheme. Starting from a low-quality reconstructed image, the objective is to learn how to suppress the artefacts due to the lack of views in the reconstructed image.

We applied and compared both strategies on two different data set. The first one is a set of 2D images obtained by generating sets of random grey-level ellipses. The second one is a set of images  of corks experimentally acquired in the context of a previous study. We evaluate the influence of multiple factors such as the geometry of acquisition, the quantity and diversity of data, or the influence of statistical noise. Both methods present quantitative improvements over a classical FBP/FDK approach, and reduce the quadratic error by an order of magnitude (from  to  for the cork data with the pre-processing method); showing that efficient CT inspection can be performed from only few dozens of images.

Presenting Author: Vo Romain CEA