High-accuracy, direct aberration determination using self-attention-armed deep convolutional neural networks

Optical microscopes have long been essential for many scientific disciplines. However, the resolution and contrast of such microscopic images are dramatically affected by aberrations. In this study, compacted with adaptive optics, we propose a machine learning technique, called the 'phase-retrieval deep convolutional neural networks (PRDCNNs)'. This aberration determination architecture is direct and exhibits high accuracy and certain generalisation ability. Notably, its performance surpasses those of similar, existing methods, with fewer fluctuations and greater robustness against noise. We anticipate future application of the proposed PRDCNNs to super-resolution microscopes.