FAN: Fourier Assignment Network for autofocus based on deep learning

IF 5 2区物理与天体物理 Q1 OPTICS Optics and Laser Technology Pub Date : 2025-02-20 DOI:10.1016/j.optlastec.2025.112579

Qizheng Liu , Xintian Mao , Jiansheng Wang , Qing Zhang , Yan Wang , Baochuan Pang , Qingli Li

{"title":"FAN: Fourier Assignment Network for autofocus based on deep learning","authors":"Qizheng Liu , Xintian Mao , Jiansheng Wang , Qing Zhang , Yan Wang , Baochuan Pang , Qingli Li","doi":"10.1016/j.optlastec.2025.112579","DOIUrl":null,"url":null,"abstract":"<div><div>AI-driven digital pathology is transforming the field, prioritizing the enhancement of image quality. In the data acquisition phase, the performance of autofocus algorithms significantly influences imaging quality. To tackle this challenge, we introduce the Fourier Assignment Network (FAN), a novel and lightweight neural network architecture aimed at predicting the defocus distance in microscopic images. FAN estimates the defocus distance using single-shot images, thus eliminating the requirement for extra hardware. Extensive experiments conducted on our dataset confirm its effectiveness, with an average error of <span><math><mrow><mn>0</mn><mo>.</mo><mn>9</mn><mspace></mspace><mi>μ</mi><mi>m</mi></mrow></math></span> across 1374 sample groups, markedly surpassing the performance of existing lightweight neural networks. Moreover, FAN is the most efficient among surveyed networks, with only 0.14G FLOPs and 0.85M parameters. The real-time and precise autofocus capability of FAN marks a notable advancement in digital pathology, providing a dependable, hardware-independent solution to enhance diagnostic accuracy.</div></div>","PeriodicalId":19511,"journal":{"name":"Optics and Laser Technology","volume":"186 ","pages":"Article 112579"},"PeriodicalIF":5.0000,"publicationDate":"2025-02-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Optics and Laser Technology","FirstCategoryId":"101","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0030399225001677","RegionNum":2,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"OPTICS","Score":null,"Total":0}

引用次数: 0

Abstract

AI-driven digital pathology is transforming the field, prioritizing the enhancement of image quality. In the data acquisition phase, the performance of autofocus algorithms significantly influences imaging quality. To tackle this challenge, we introduce the Fourier Assignment Network (FAN), a novel and lightweight neural network architecture aimed at predicting the defocus distance in microscopic images. FAN estimates the defocus distance using single-shot images, thus eliminating the requirement for extra hardware. Extensive experiments conducted on our dataset confirm its effectiveness, with an average error of

0.9 μ m

across 1374 sample groups, markedly surpassing the performance of existing lightweight neural networks. Moreover, FAN is the most efficient among surveyed networks, with only 0.14G FLOPs and 0.85M parameters. The real-time and precise autofocus capability of FAN marks a notable advancement in digital pathology, providing a dependable, hardware-independent solution to enhance diagnostic accuracy.

查看原文

微信好友朋友圈 QQ好友复制链接

本刊更多论文

FAN：基于深度学习的自动对焦傅立叶分配网络

人工智能驱动的数字病理学正在改变这一领域，优先考虑提高图像质量。在数据采集阶段，自动对焦算法的性能直接影响成像质量。为了解决这一挑战，我们引入了傅里叶分配网络（FAN），这是一种新颖的轻量级神经网络架构，旨在预测微观图像中的离焦距离。FAN使用单张图像估计离焦距离，从而消除了对额外硬件的需求。在我们的数据集上进行的大量实验证实了它的有效性，在1374个样本组中平均误差为0.9μm，明显超过了现有的轻量级神经网络的性能。此外，FAN是被调查网络中效率最高的，只有0.14G的FLOPs和0.85M的参数。FAN的实时和精确的自动对焦能力标志着数字病理学的显著进步，提供了一个可靠的，独立于硬件的解决方案，以提高诊断的准确性。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文去求助

来源期刊

Optics and Laser Technology 物理-光学

CiteScore

8.50

自引率

10.00%

发文量

1060

审稿时长

3.4 months

期刊介绍： Optics & Laser Technology aims to provide a vehicle for the publication of a broad range of high quality research and review papers in those fields of scientific and engineering research appertaining to the development and application of the technology of optics and lasers. Papers describing original work in these areas are submitted to rigorous refereeing prior to acceptance for publication. The scope of Optics & Laser Technology encompasses, but is not restricted to, the following areas: •development in all types of lasers •developments in optoelectronic devices and photonics •developments in new photonics and optical concepts •developments in conventional optics, optical instruments and components •techniques of optical metrology, including interferometry and optical fibre sensors •LIDAR and other non-contact optical measurement techniques, including optical methods in heat and fluid flow •applications of lasers to materials processing, optical NDT display (including holography) and optical communication •research and development in the field of laser safety including studies of hazards resulting from the applications of lasers (laser safety, hazards of laser fume) •developments in optical computing and optical information processing •developments in new optical materials •developments in new optical characterization methods and techniques •developments in quantum optics •developments in light assisted micro and nanofabrication methods and techniques •developments in nanophotonics and biophotonics •developments in imaging processing and systems