Modeling Uncertainty in Deep Learning for Focus Prediction of Defocused Images

Autofocus is a critical component in automatic digital microscopy, directly impacting imaging precision and operational efficiency. To accelerate focusing processes, deep learning methods are increasingly utilized for single-frame focus prediction. However, conventional networks often assume their predictions are accurate across all sample types, which risks undetected errors when handling novel or heterogeneous samples. This paper introduces a Bayesian convolutional neural network specifically designed to predict defocus distance from single images while simultaneously generating uncertainty estimates for these predictions. Leveraging these uncertainty estimates, our method dynamically assesses and filters the reliability of the focus predictions, ensuring robust performance across variable sample types. Through comprehensive experiments on diverse specimens, including pathological tissue sections and printed circuit board inspections, our approach demonstrates superior prediction accuracy and efficiency compared to traditional data-driven methods and sharpness quantification functions. Furthermore, the uncertainty estimation effectively flags unreliable predictions, substantially enhancing autofocus reliability. Based on this uncertainty framework, we additionally propose an innovative focal plane alignment strategy, further advancing the system's applicability in complex imaging environments.