Learning to high-performance autofocus microscopy with laser illumination
Zhijie Hua, Xu Zhang, Dawei Tu, Xingzhao Wang, Nuodi Huang
Abstract
The autofocusing plays a crucial role in various industrial microscopic measurements. However, conventional autofocus methods involve axial scanning and are prone to inaccuracies caused by surface noises, such as defects and texture, making them both time-consuming and unreliable. In this work, a novel high-performance autofocus pipeline with laser illumination was introduced, namely Learning Split-image Prediction Model (LSPM), to accurately predict the defocus distance from a single split image. To effectively reduce the interference of the noises as foregrounds, a lightweight generation network was built to restore the incomplete split images. Furthermore, a new judgment mechanism for split images was proposed to identify the ideal focal plane of each focal stack. The experiment demonstrated LSPM outperforms the state-of-the-art methods, in terms of focusing accuracy, especially when the defocusing amount is relatively small, achieving only 1/15 depth of field. LSPM has the advantages of high accuracy, high speed, and robustness in industrial measurement.