Michał Gontarz
supervisor: Małgorzata Kujawińska
Whenever we perform a measurement with the use of interferometric or grid methods, we want to obtain phase information about a measured object. 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 the complexity of discontinuities, solving this problem with traditional image processing for real phase images is time consuming and unreliable.
Therefore I propose a CNN based pipeline, which consists of a denosing step, done by a small U-Net architecture CNN, and an unwrapping step proposed in two ways: semantic segmentation based technique with an Attention U-Net architecture and image translation with a U-Net architecture with residual blocks and a leaky-ReLU activation function.
The noise removal has been able to remove noise and retain crucial information about the phase distribution on computer generated phase images, as well as real phase images obtained from holographic tomography. Each phase unwrapping pipeline direction has shown good performance and little unwrapping error, whilst, in some cases, outperforming conventional unwrapping methods. The pipeline shows great promise and high accuracy even with a difficult wrapped phase distributions obtained from optical diffraction tomography and optical coherence tomography measurements of highly diffusive objects.