Processing speckle patterns with model-trained neural networks

Artificial neural networks can be used to process patterns corrupted by the laser speckle effect. This paper discusses an examples where neural networks were used to detect structural damage using characteristic fringe patterns as input. The artificial neural networks were trained with fringe patterns generated from a finite element model and a simple model of the laser speckle effect. The neural networks were tested on patterns generated by both models and real structures. The neural networks are being developed as high-speed processors for electronic holography. This paper quantifies the overhead required to make neural networks robust to the laser speckle effect. There is a discussion of the ability of these networks to generalize at finite element resolutions on the underlying fringe patterns. The ultimate objective is to test whether combinations of electronic holography and neural networks can be effective interfaces between computational models and experiments or tests.