Denoisers in lensless single-shot super-resolution phase retrieval

Abstract

The goal of this thesis is to analyse the performance of denoisers in a lensless Single-shot Super-Resolution Phase Retrieval (SSR-PR) algorithm with both continuous and binary target images. Specifically, we will analyse the performance of SSR-PR with the use of a Block Matching 3D filter (BM3D) (belonging to the class of model-based denoisers) and DRUNet (neural network-based denoiser). Two variations of DRUNet are studied. The first variation assumes the noise model to be additive white Gaussian noise and takes a single parameter of the noise variance as an input together with a noisy image. The second variation assumes the spatially varying noise model and uses a neural network-based method called PIXPNet to estimate the noise map, which is used as an auxiliary input. We refer to this second variation of DRUNet as the Blind denoiser as it calculates the noise map independently. In this thesis, two phase retrieval algorithms are introduced. These are the state-of-the-art SSR-PR algorithm and the SR-SPAR algorithm on which it is based. Additionally, the required background information on the methods is provided. In phase retrieval, the goal is to estimate the phase characteristics of the object being imaged based on amplitude-only measurement of the coherent light diffracted by the object. This process is difficult since traditional sensors cannot measure the phase of the light, and there are many sources of possible noise and errors in the imaging systems. The two methods utilize a phase mask, which diffracts the light in a coded pattern across the sensor to acquire additional information to reconstruct the phase characteristics. This phase mask also replaces the lenses in the system, making it more compact and cheaper to manufacture. However, this phase mask causes additional errors, which make reconstruction of the scene more challenging. Additionally, the algorithm used for the analysis in this thesis uses only one image of the complex scene, making the phase reconstruction even more difficult and prone to noise. To suppress these noises and errors, efficient filtering can be used to achieve better quality reconstructions of the object. In addition, the algorithm was used for super-resolution reconstruction of the scenes, giving a higher pixel-wise resolution. The performance of the denoisers was analysed using a script that simulates a physical phase retrieval imaging system using a coherent illumination source to illuminate the object being captured, a phase mask, and a sensor to record the diffracted wavefront of the object. The denoisers' performance was tested in the ideal noiseless case and with various levels of added noise for both the continuous and binary targets. Through the testing, we found that the denoisers perform very differently depending on the amount of noise present and the object being imaged. BM3D and the normal DRUNet performed well in the lower noise cases. The Blind denoiser performed best with noisy continuous data, while DRUNet performed better than it with noisy binary data. DRUNet and the Blind denoiser had issues with artifacts and disappearing information with added noise, while the BM3D reconstructions were heavily corrupted by the noise.