In image reconstruction, we usually design reconstruction algorithms in such a way that we can choose the most convinent sampling pattern. However, the choice of the sampling pattern has a significant inpact on the final quality. In compressed sensing (CS) we noticed the importance of using incoherent sampling in order to have successful sparse reconstructions, which seems to me a good example of how sampling and reconstruciton should be matched. However, machine learning and deep learning have significantly changed the way image reconstruction is done. So, how should the sampling pattern be designed?
Let me know your opinion about it?
Marcelo V. W. Zibetti interesting question, for our application (diffuse optical tomography) a key challenge is recognizing the often fatal heterogeneity of cancer lesions, where we are not that interested in healthy tissue, so a learned sampling that adapts to marked variance in optical coefficients of abnormal tissue would be ideal, but this is often a circular problem (how to know the coefficients without recon). A multipass/stage approach could work around this, ideally we'd prefer focussing maximum information (and thus compute/learning) on relatively sparse regions yet without risking overfitting or introduction of artifacts.
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude