Abstract
We propose an untrained deep neural network framework for joint amplitude–phase retrieval in single-pixel imaging (SPI). In many advanced optical applications, e.g., quantitative phase microscopy and coherent diffractive imaging, the object of interest is inherently complex-valued. However, conventional SPI approaches are designed for amplitude-only reconstruction, and the joint recovery of both amplitude and phase from intensity-only measurements is a fundamentally non-linear and severely ill-posed problem. In the proposed method, the unknown complex object is parameterized as the output of a single U-Net, whose weights are optimized directly to fit the observed SPI measurements without any training data. The deep image prior paradigm embedded in the convolutional architecture serves as an effective regularizer. Numerical simulations on 128×128 complex-valued objects with a sampling ratio of 400% demonstrate that the proposed method achieves high reconstruction accuracy. It is believed that the proposed method provides a practical and training-data-free solution for complex-field SPI.