Phase Retrieval Using Conditional Generative Adversarial Networks

Abstract

In this paper, we propose the application of conditional generative adversarial networks to solve various phase retrieval problems. We show that including knowledge of the measurement process at training time leads to an optimization at test time that is more robust to initialization than existing approaches involving generative models. In addition, conditioning the generator network on the measurements enables us to achieve much more detailed results. We empirically demonstrate that these advantages provide meaningful solutions to the Fourier and the compressive phase retrieval problem and that our method outperforms well-established projection-based methods as well as existing methods that are based on neural networks. Like other deep learning methods, our approach is very robust to noise and can therefore be very useful for real-world applications.

DeepPhaseCut: Deep Relaxation in Phase for Unsupervised Fourier Phase Retrieval

“…In the second category, a network is trained in an end-to-end manner to restore the signal from the given measured magnitude through supervised learning scheme [22], [23]. For example, for a large number of paired magnitude and label data set (b, x), the neural network Q Θ parameterized by Θ is trained by minimizing the l 2 loss:…”

“…For example, in [22], a hybrid use of two networks and the HIO method was proposed to address the phase retrieval problem. Uelwer et al [23] proposed using the Fourier measurement and the latent variable together as input to train the conditional generative adversarial network (GAN) [32].…”

“…We changed the RGB images to grayscale images, and resized them from the center cropped images of 108 × 108 pixels to 64 × 64 pixels. This preprocessing was identical to that employed in [20], [23]. Among 202,599 images, 199,000 images were used for training, 1,121 images for validation, and the rest for testing.…”

“…Recently, various forms of deep learning approaches have been investigated as a fast and effective approach for phase retrieval [19], [20], [21], [22], [23], [24], [25], [26]. All of these algorithms have demonstrated high performance compared to the traditional alternating projection-based methods.…”

“…All of these algorithms have demonstrated high performance compared to the traditional alternating projection-based methods. Nonetheless, iterative schemes using a pretrained network as a regularization term [19], [20] leads to increased computational complexity, whereas feedforward methods usually require extensive amount of training data and are often sensitive to the inherent ambiguity of the solutions [22], [23], [25], [26].…”