Verbeterde reconstructie van snelle CT-scans op basis van a priori informatie

An industrial computed tomography (CT) system is applied for defect detection and dimensional analysis for internal features of manufactured parts or assemblies. In-line CT is a non-destructive three-dimensional (3-D) imaging technology integrated directly into production lines for quality control. The process of quality control using CT begins with acquiring two-dimensional (2-D) X-ray projections from various angles of a rotating object, usually spanning 0 to 360 degrees. These projections are recorded by a detector panel, where each image demonstrates the attenuation of X-ray intensity resulting from the X-rays' interaction with the atoms of the object. Applying dedicated algorithms, a 3-D volume is reconstructed from the 2-D projections and subsequently analysed for quality control, e.g., CT metrology. To accelerate the scanning process and increase throughput in production, reducing the number of projections through sparser angular sampling is an efficient approach. However, this results in under-sampled CT data or few-view CT data, which compromises the quality of reconstruction and accuracy of the CT metrology. This manuscript addresses the challenge by enhancing the reliability and precision of fast scanning strategies through exploiting a priori knowledge and developing dedicated reconstruction algorithms for in-line CT. Three reconstruction algorithms based on total variation (TV) minimization are proposed. To overcome the limitation inherent to the TV-based method, which may remove some small features during reconstruction, a novel technique for estimating the noise level is introduced. The estimated noise level is incorporated into the TV-type method using two distinct strategies. In the first, the noise level estimation supports the automation of TV-type reconstruction by selecting the optimal regularization parameter according to the noise level, leading to the unconstrained split Bregman algorithm (USB-TV). In the second strategy, the noise level serves as a constraint within the Bregman iterative procedure, enabling the "error forgetting" feature. This yields the constrained split Bregman algorithm (CSB-TV), enhancing image contrast and metrology. The third developed algorithm utilizes infimal convolution regularization to incorporate a priori knowledge. In this model, the reconstruction is segmented into two main components: structure and pores, penalized by TV-type and ℓ1-norm minimization, respectively. The final volume is obtained by summing these two components. The effectiveness of the proposed methods is evaluated against full-view reconstructions obtained using standard techniques such as the Feldkamp-Davis-Kress (FDK) algorithm. Throughout all experiments, using only between 2%–12% of data required by FDK, the proposed algorithms yielded a higher signal-to-noise ratio with metrology results comparable to those of full-view FDK.