In white-light diffraction phase imaging, when used with insufficient spatial filtering, phase image exhibits object-dependent artifacts, especially around the edges of the object, referred to the well-known halo effect. Here we present a new deep-learning-based approach for recovering halo-free white-light diffraction phase images. The neural network-based method can accurately and rapidly remove the halo artifacts not relying on any priori knowledge. First, the neural network, namely HFDNN (deep neural network for halo free), is designed. Then, the HFDNN is trained by using pairs of the measured phase images, acquired by white-light diffraction phase imaging system, and the true phase images. After the training, the HFDNN takes a measured phase image as input to rapidly correct the halo artifacts and reconstruct an accurate halo-free phase image. We validate the effectiveness and the robustness of the method by correcting the phase images on various samples, including standard polystyrene beads, living red blood cells and monascus spores and hyphaes. In contrast to the existing halo-free methods, the proposed HFDNN method does not rely on the hardware design or does not need iterative computations, providing a new avenue to all halo-free white-light phase imaging techniques.
Quantitative phase imaging in dual-wavelength interferometry using a single wavelength illumination and deep learning
