Learning-based phase imaging using a low-bit-depth pattern

Phase imaging always deals with the problem of phase invisibility when capturing objects with existing light sensors. However, there is a demand for multiplane full intensity measurements and iterative propagation process or reliance on reference in most conventional approaches. In this paper, we present an end-to-end compressible phase imaging method based on deep neural networks, which can implement phase estimation using only binary measurements. A thin diffuser as a preprocessor is placed in front of the image sensor to implicitly encode the incoming wavefront information into the distortion and local variation of the generated speckles. Through the trained network, the phase profile of the object can be extracted from the discrete grains distributed in the low-bit-depth pattern. Our experiments demonstrate the faithful reconstruction with reasonable quality utilizing a single binary pattern and verify the high redundancy of the information in the intensity measurement for phase recovery. In addition to the advantages of efficiency and simplicity compared to now available imaging methods, our model provides significant compressibility for imaging data and can therefore facilitate the low-cost detection and efficient data transmission.