首页|基于深度学习的微藻自动检测系统研究

基于深度学习的微藻自动检测系统研究

扫码查看
目的 微藻养殖产业规模巨大,在养殖过程中微藻易受杂菌和其他污染物的影响,因此需要定期对微藻进行检测,以确定其生长情况。现有的光学显微成像法和光谱分析法对实验人员、实验设备及场地的要求较高,无法做到实时快速检测。为了实现实时快速检测,需要一套检测要求低、速度快的实时微藻检测系统。方法 本文开发了一种基于深度学习的微藻检测系统,通过搭建一套基于明场成像的显微成像设备,使用采集的图像训练基于YOLOv3的神经网络,并将训练好的神经网络部署到微型计算机,从而实现了实时便携微藻检测。本文对特征提取网络进行改进,包括引入跨区域残差连接机制和注意力选择机制,另外还将优化器改为Adam优化器,使用多阶段多方法组合策略。结果 加载跨区域残差连接机制时最高平均精度(mAP)值为0。92。通过与人工结果进行对比,得到检测误差为2。47%。结论 该系统能够实现微藻实时便携检测,提供较为准确的检测结果,可以应用于微藻养殖中的定期检测。
Research on Automatic Microalgae Detection System Based on Deep Learning
Objective The scale of microalgae farming industry is huge.During farming,it is easy for microalgae to be affected by miscellaneous bacteria and other contaminants.Because of that,periodic test is necessary to ensure the growth of microalgae.Present microscopy imaging and spectral analysis methods have higher requirements for experiment personnel,equipment and sites,for which it is unable to achieve real-time portable detection.For the purpose of real-time portable microalgae detection,a real-time microalgae detection system of low detection requirement and fast detection speed is needed.Methods This study has developed a microalgae detection system based on deep learning.A microscopy imaging device based on bright field was constructed.With imaged captured from the device,a neural network based on YOLOv3 was trained and deployed on microcomputer,thus realizing real-time portable microalgae detection.This study has also improved the feature extraction network by introducing cross-region residual connection and attention mechanism and replacing optimizer with Adam optimizer using multistage and multimethod strategy.Results With cross-region residual connection,the mAP value reached 0.92.Compared with manual result,the detection error was 2.47%.Conclusion The system could achieve real-time portable microalgae detection and provide relatively accurate detection result,so it can be applied to periodic test in microalgae farming.

microalgae detectionbrightfield microscopydeep learningobject detection

向睿捷、刘浩、路珍、肖泽宇、刘海鹏、王寅初、彭晓、严伟

展开 >

深圳大学物理与光电工程学院,深圳市光子学与生物光子学重点实验室,光电子器件与系统教育部/广东省重点实验室,深圳 518060

中国科学院烟台海岸带研究所,烟台 264003

国家基础学科公共科学数据中心,北京 100190

微藻检测术 明场显微术 深度学习 目标识别

国家重点基础研究发展计划国家自然科学基金国家自然科学基金国家自然科学基金国家自然科学基金广东省高等学校科技创新(重点)项目广东省基础与应用基础研究项目广东省基础与应用基础研究项目深圳市科技研究项目

2021YFF0502900619751273177158461835009422061442021ZDZX20132022A15150119542022A1515011845JCYJ20220531102807017

2024

生物化学与生物物理进展
中国科学院生物物理研究所,中国生物物理学会

生物化学与生物物理进展

CSTPCD北大核心
影响因子:0.476
ISSN:1000-3282
年,卷(期):2024.51(1)
  • 46