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

Paper Number: 98042

98042 - Sparse-View X-Ray Ct Reconstruction Using Cad Model Registration 

X-ray Computed Tomography is a powerful non-destructive testing tool increasingly used by manufacturers for ensuring the conformity of the produced parts. Despite a growing interest, it is struggling to establish itself in online testing applications due to the large number of X-ray projections it requires to ensure a good reconstructed image. To reduce this number of projections from few hundreds to few dozens while still getting good reconstruction quality, we propose to infer a so-called ‘mask’ on the volume to be reconstructed. By constraining the back-projection of the acquired X-ray projections only on this mask, corresponding to the voxels of the volume containing matter, iterative reconstruction algorithms, already very efficient at a low number of views compared to the traditional FDK, can better reconstruct an object, and with fewer computational resources. However, this technique requires a preliminary step: the registration of the experimental data to the a priori mask data.

In this work, we consider the CAD model of the object as a mask used to constrain the reconstruction. In a first step, we perform the CAD model registration by virtually moving this 3D geometric model in a simple radiographic projection simulator until the simulated projection matches the experimental one, using an algorithm inspired by Iterative Inverse Perspective Matching. Results obtained on both simulated and experimental images demonstrate the robustness of the method, which ensures an accurate registration in few seconds even in the case of slight differences between the real object and its theoretical CAD model.

In a second step, and once the registration process has been completed, we create a mask adapted to our object using the registered CAD model. Only voxels included in this mask are considered as unknown, which greatly reduces the dimension of the problem to solve, especially in the case of hollow shapes. Based on the registered mask, we achieve good reconstruction results from no more than a dozen of projections. This approach, which is validated on realistic images simulated with the radiographic module of CIVA, ensures the detection of inserted flaws not visible with a classical iterative approach. The results show typically the potential of such approach for on-line static systems of less than 12 views.

Presenting Author: Victor Bussy Université Paris-Saclay, CEA, List