当前位置: X-MOL 学术Opt. Laser Technol. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Automatic classification of marine plankton with digital holography using convolutional neural network
Optics & Laser Technology ( IF 5 ) Pub Date : 2021-02-16 , DOI: 10.1016/j.optlastec.2021.106979
Yilong Zhang , Yaoxiang Lu , Haixia Wang , Peng Chen , Ronghua Liang

The observation, statistics and classification of marine plankton are the basis of marine ecological research. In recent years, digital holography technology has been widely used in the detection of plankton, living cells and particles. The digital holographic microscope can record the three-dimensional information of the sample in a non-contact manner, which has great advantages for the observation of marine plankton. However, the reconstruction of holographic images requires a priori knowledge and time-consuming iterative calculations, which affects the application efficiency of digital holography. In order to solve the above problem, we proposed to directly predict marine plankton species by training a convolutional neural network (CNN) to extract information of raw holograms without performing reconstruction. In our method, machine vision is used to directly identify the original hologram, and the reconstruction process is omitted, thus reducing time consumption. By using deep transfer learning (DTL), neural networks are trained without starting from scratch and the requirements on training data and training time are alleviated. The classification accuracy of our method could reach approximately 92% for experimental datasets. Our approach demonstrated efficient and robust performance for raw plankton holograms processing.



中文翻译:

使用卷积神经网络的数字全息自动分类海洋浮游生物

海洋浮游生物的观测,统计和分类是海洋生态研究的基础。近年来,数字全息技术已广泛用于检测浮游生物,活细胞和微粒。数字全息显微镜可以非接触方式记录样品的三维信息,对于海洋浮游生物的观测具有很大的优势。然而,全息图像的重建需要先验知识和费时的迭代计算,这影响了数字全息的应用效率。为了解决上述问题,我们建议通过训练卷积神经网络(CNN)直接提取海洋全息图信息,而无需进行重建,直接预测海洋浮游生物的种类。用我们的方法 机器视觉用于直接识别原始全息图,并且省略了重建过程,从而减少了时间消耗。通过使用深度迁移学习(DTL),无需从头开始即可对神经网络进行训练,从而减轻了对训练数据和训练时间的要求。对于实验数据集,我们方法的分类精度可以达到约92%。我们的方法证明了原始浮游生物全息图处理的高效和鲁棒性。对于实验数据集,我们方法的分类精度可以达到约92%。我们的方法证明了原始浮游生物全息图处理的高效和鲁棒性。对于实验数据集,我们方法的分类精度可以达到约92%。我们的方法证明了原始浮游生物全息图处理的高效和鲁棒性。

更新日期:2021-02-17
down
wechat
bug