In this thesis, I investigate how much critical decision-making information survives when an industrial computed tomography (CT) scanner acquires only a sparse set of projections, and to what extent a neural network can recover it. For industrial applications, such as inspecting a log before sawing or checking a sealed jar on a production line, the reconstructed image is never an end in itself. It exists solely to support an automated decision. Traditional high-fidelity reconstruction requires dense angular sampling, demanding expensive and bulky hardware. Reducing the number of projections solves these practical constraints but renders the mathematical inversion severely ill-posed, overwhelming classical algorithms with artefacts. While standard literature evaluates sparse-view reconstruction by measuring how closely the output visually resembles a high-resolution reference, this work takes a different approach. It measures success by whether the specific physical quantities required for the industrial task can still be reliably extracted. The thesis is divided into two parts. The first establishes the theoretical foundation of the research. It covers the physics of X-ray attenuation and the mathematics of tomographic reconstruction, from the Radon transform to filtered backprojection, explicitly defining the sampling limits that sparse acquisitions violate. It then outlines the deep learning architectures used in this work before concluding with a literature review of sparse-view reconstruction and the methodological gaps it leaves regarding task-oriented industrial inspection. The second part presents two distinct industrial applications developed during an internship at Microtec SpA. Both are driven by a unified supervised learning pipeline, in which a network is trained on paired sparse and high-resolution reconstructions, but is explicitly forced to predict the precise structural parameters the application consumes alongside the reconstructed volume itself. The first application tackles the reconstruction of wooden logs using just twenty-five projections acquired from a single side. The network processes this severely limited reconstruction alongside an externally measured bark contour. Beyond simply enhancing the image, the model predicts the coordinates of the pith and the underbark contour used to calculate the log's grading diameter. This predicted diameter achieves a mean absolute error of just 2.31 mm on logs averaging 259 mm. Furthermore, this pipeline is used as an empirical tool to compare competing acquisition geometries, testing the limits of recovery down to an extreme two-view configuration. The second application addresses the inspection of sealed glass containers for physical contaminants, a challenging task where the target is both unpredictably located and only a few pixels wide. In this scenario, I compare two distinct strategies for reducing the acquisition to one fifth of the standard projections. Both strategies yield reconstructions that standard image metrics rank as equivalent, with a difference of 0.3 dB in PSNR in favour of the second. On the detectability of the contaminant the ranking is reversed, and by a wide margin: the area under the precision-recall curve is 0.526 for the first strategy against 0.086 for the second. Global image quality therefore does not rank the two reductions the way the inspection does.

A Deep Learning Approach to Ultra-Sparse-View CT Reconstruction: Applications in the Wood and Food Industries

ZORZI, FRANCESCA
2025/2026

Abstract

In this thesis, I investigate how much critical decision-making information survives when an industrial computed tomography (CT) scanner acquires only a sparse set of projections, and to what extent a neural network can recover it. For industrial applications, such as inspecting a log before sawing or checking a sealed jar on a production line, the reconstructed image is never an end in itself. It exists solely to support an automated decision. Traditional high-fidelity reconstruction requires dense angular sampling, demanding expensive and bulky hardware. Reducing the number of projections solves these practical constraints but renders the mathematical inversion severely ill-posed, overwhelming classical algorithms with artefacts. While standard literature evaluates sparse-view reconstruction by measuring how closely the output visually resembles a high-resolution reference, this work takes a different approach. It measures success by whether the specific physical quantities required for the industrial task can still be reliably extracted. The thesis is divided into two parts. The first establishes the theoretical foundation of the research. It covers the physics of X-ray attenuation and the mathematics of tomographic reconstruction, from the Radon transform to filtered backprojection, explicitly defining the sampling limits that sparse acquisitions violate. It then outlines the deep learning architectures used in this work before concluding with a literature review of sparse-view reconstruction and the methodological gaps it leaves regarding task-oriented industrial inspection. The second part presents two distinct industrial applications developed during an internship at Microtec SpA. Both are driven by a unified supervised learning pipeline, in which a network is trained on paired sparse and high-resolution reconstructions, but is explicitly forced to predict the precise structural parameters the application consumes alongside the reconstructed volume itself. The first application tackles the reconstruction of wooden logs using just twenty-five projections acquired from a single side. The network processes this severely limited reconstruction alongside an externally measured bark contour. Beyond simply enhancing the image, the model predicts the coordinates of the pith and the underbark contour used to calculate the log's grading diameter. This predicted diameter achieves a mean absolute error of just 2.31 mm on logs averaging 259 mm. Furthermore, this pipeline is used as an empirical tool to compare competing acquisition geometries, testing the limits of recovery down to an extreme two-view configuration. The second application addresses the inspection of sealed glass containers for physical contaminants, a challenging task where the target is both unpredictably located and only a few pixels wide. In this scenario, I compare two distinct strategies for reducing the acquisition to one fifth of the standard projections. Both strategies yield reconstructions that standard image metrics rank as equivalent, with a difference of 0.3 dB in PSNR in favour of the second. On the detectability of the contaminant the ranking is reversed, and by a wide margin: the area under the precision-recall curve is 0.526 for the first strategy against 0.086 for the second. Global image quality therefore does not rank the two reductions the way the inspection does.
2025
A Deep Learning Approach to Ultra-Sparse-View CT Reconstruction: Applications in the Wood and Food Industries
Computed Tomography
Deep Learning
Image Reconstruction
File in questo prodotto:
File Dimensione Formato  
Zorzi_Francesca.pdf

Accesso riservato

Dimensione 9.2 MB
Formato Adobe PDF
9.2 MB Adobe PDF

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/118559