Hoppa till huvudinnehållet

Post-doc position: Deep learning based 3D CT reconstruction in forestry

The position is part of a collaboration between the Mathematical Imaging Sciences group within the Department of Mathematics at KTH - Royal Institute of Technology and the CT Wood Centre within the Division of Wood Science and Engineering at Luleå university of technology.

The overall goal is to develop theory and algorithms for image-guided optimisation of the sawline in forestry. One research topic is to develop and implement deep neural networks suited for tomographic 3D image reconstruction. This involves designing physics aware network architectures that incorporate an explicit physics model for x-ray imaging, like those obtained by unrolling a suitable iterative scheme. A related topic is to perform reconstruction with uncertain acquisition geometry, i.e., performing 3D reconstruction while recovering data acquisition parameters. A final topic is to combine such physics aware (deep) neural networks for reconstruction with neural network(s) for downstream post-processing, like semantic 3D segmentation of interior anomalies (knots, cracks, etc) in logs or networks for extracting of a 3D orientation field (wooden fibers) of a log. As shown in this paper, one can combine reconstruction with post-processing and after re-training, one obtains a deep neural network for joint 3D image reconstruction and post-processing.

The implementation of the above neural networks is challenging due to the time-critical nature of the use case. In particular, one needs to manage memory footprint and computational feasibility is an issue during training. Prototypes will be implemented as components in ODL which preferably utilises the connection to frameworks for tomography (ASTRA) and deep learning (PyTorch). 

 


Profilbild av Ozan Öktem

Portfolio

  • Post-doc position: Deep learning based 3D CT reconstruction in forestry