Statistically guided improvements in speckle phase discontinuity predictions by machine learning systems

Abstract

Abstract. Before a machine learning system can be deployed, it typically needs to undergo a training phase, which enables it to acquire the necessary structures and information to solve similar problems. The performance of a machine learning system is commonly assessed by measuring how well the system is able to solve problems, which are generally similar but not identical to those used as examples during the training phase. We investigate applying machine learning to a key step of interferogram analysis, namely the identification of phase discontinuities. Identifying phase discontinuities correctly can be an especially challenging task due to the inherently noisy nature of speckle interferograms. Traditionally, automated edge detection operators are employed for this task, often producing inferior results as compared to those produced manually by an expert human operator. We present a machine learning approach to the phase discontinuity identification problem, discuss its potential and merits, and examine the challenges encountered during the training phase. We describe novel measures for quantifying the learning attainment levels of the system and describe how these measures can be used to guide the training phase in a methodical and intuitive manner.