Michal Gontarz
supervisor: Małgorzata Kujawińska
Whenever we perform a measurement with the use of interferometric or grid methods, we aim in obtaining phase information in which the object’s shape, displacement or other measurand is coded. However, due to the nature of phase retrieval algorithms, the results are wrapped into phase mod(2π). In order to obtain a continuous phase distribution, the phase discontinuities have to be eliminated in a process known as unwrapping. Due to significant noise and a complexity of discontinuities in real phase images, solving this problem with traditional image processing is time consuming and unreliable.
Therefore I propose a CNN based pipeline, which consists of two steps: denoising of wrapped phase images and eliminating discontinuities. The denoising is done with a small encoder-decoder architecture CNN (U-Net). Unwrapping is solved by semantic segmentation of wrapped phase images, where one label corresponds to one phase wrap level. It is done by a similar architecture, however, bigger and more complex.
The performance of this pipeline was tested on computer generated phase images with complex distributions and on realistic phase images, which originate from holographic tomography measurements. In both cases denoising has been successful, whilst not sacrificing information at phase discontinuities, whilst unwrapping is done efficiently and reliably. With both models of CNN working together, both noisy synthetic and experimental phase fringes have been unwrapped with high accuracy.