Automated Focus Distance Estimation for Digital Microscopy Using Deep Convolutional Neural Networks

Tathagato Rai Dastidar; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

Abstract


An essential component of an automated digital microscopy system is auto focusing, which involves moving the microscope stage along the vertical axis to find the position where the underlying image is the sharpest. Auto focusing algorithms deployed in current commercially available digital microscopes cannot match the efficiency of a trained human operator. Traditionally, auto focusing has been achieved by acquiring multiple images in the vertical direction and maximising a measure of image sharpness. This paper presents a method for auto focusing based on deep convolutional neural networks (CNN). Given two images in the vertical focus stack, the CNN predicts the optimal distance the stage needs to be moved to achieve best focus, relative to the current position. The method was trained and results are demonstrated on a publicly available data set. It is shown to outperform previously published work on this data set. The compute and memory requirements of the model are shown to be ideal for deployment in an edge device with limited computing resources.

Related Material


[pdf]
[bibtex]
@InProceedings{Dastidar_2019_CVPR_Workshops,
author = {Rai Dastidar, Tathagato},
title = {Automated Focus Distance Estimation for Digital Microscopy Using Deep Convolutional Neural Networks},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}