A coarse-to-fine attention-guided autofocusing for holography under high noisy scenes with explainable neural network

Abstract

Autofocusing of digital holography typically relies on various criterion functions to evaluate the focus quality. However, these functions often struggle to accurately determine the focal plane in highly interferential environments. In this paper, we present a method for autofocusing of low-quality digital holographic images under extreme high noise environments. This approach incorporates a neural network as part of the solution but does not rely solely on its output, thereby overcoming uncertainties during the computation process. A key feature of our approach is the application of neural network attention mechanisms. These mechanisms excel at recognizing key areas within an image that significantly impact focus quality, thereby enabling precise focus metric calculations in complex visual settings. In our approach, the design of neural network relies solely on distinguishing focused from unfocused areas, a relatively simple task for neural networks. This design reduces our dependency on large datasets. Additionally, due to its modular construction, our method can be effortlessly integrated into diverse imaging contexts, demonstrating excellent plug-and-play capabilities. Experimental results demonstrate that our method not only enhances the precision of autofocus in digital holography but also shows promise in extending its applicability to other scientific and engineering fields. Our findings suggest potential for the broader application of deep learning in addressing analogous challenges in image analysis, opening new avenues for intelligent, data-efficient image processing.