Discrimination

Gordon, Mirta B.

doi:10.1007/3-540-28847-3_6

Cited by 1 publication

(1 citation statement)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There are 100 input nodes, which naturally matches the number of pixels in the 10 × 10 pixel image sections on which the network operates. For this work, we used the hyperbolic tangent (tanh) as the sigmoid activation function [39], which offers some practical advantages [40] over the logistic sigmoid [41] function.…”

Section: A Neural Network Employedmentioning

confidence: 99%

Phase discontinuity predictions using a machine-learning trained kernel

Sawaf

Groves

2014

Appl. Opt.

View full text Add to dashboard Cite

Phase unwrapping is one of the key steps of interferogram analysis, and its accuracy relies primarily on the correct identification of phase discontinuities. This can be especially challenging for inherently noisy phase fields, such as those produced through shearography and other speckle-based interferometry techniques. We showed in a recent work how a relatively small 10 × 10 pixel kernel was trained, through machine learning methods, for predicting the locations of phase discontinuities within noisy wrapped phase maps. We describe here how this kernel can be applied in a slidingwindow fashion, such that each pixel undergoes 100 phase-discontinuity examinations-one test for each of its possible positions relative to its neighbors within the kernel's extent. We explore how the resulting predictions can be accumulated, and aggregated through a voting system, and demonstrate that the reliability of this method outperforms processing the image by segmenting it into more conventional 10 × 10 nonoverlapping tiles. When used in this way, we demonstrate that our 10 × 10 pixel kernel is large enough for effective processing of full-field interferograms. Avoiding, thus, the need for substantially more formidable computational resources which otherwise would have been necessary for training a kernel of a significantly larger size.

show abstract

Section: A Neural Network Employedmentioning

confidence: 99%