Autonomous Plankton Classification from Reconstructed Holographic Imagery by L1- PCA-assisted Convolutional Neural Networks

Kavita Varma, Lisa Nyman, Konstantinos Tountas, George Sklivanitis, A. R. Nayak, D. Pados

2020 Global Oceans 2020: Singapore – U.S. Gulf Coast Cited 4 times

Abstract

Studying and monitoring plankton distribution is vital for global climate and environment protection as they are the most elementary part of oceanic eco-systems. However, the conventional methods and techniques used for understanding the planktons are slow and lacks precision and therefore, in modern day scientific and engineering implementations, Convolutional Neural Networks is extensively used in deep learning and machine learning applications as it outperforms traditional manual approach. Dynamic nature of oceans make it very challenging to monitor these microscopic organisms. Our approach here is to generate a powerful automated plankton recognition system to autonomously identify them and improve the D-CNN for classification of the Plankton holographic imagery curated with the method of Data Conformity Evaluation. The performance of D-CNN classifier is improved by various hyper-parameter tuning, regularization techniques and appending meta-data. Conformity evaluation is based on a matric that's calculated on a continuously refined sequence of calculated $L_{1}$-norm tensor subspaces of the Plankton images. We note that our classifier performs accurately where our results improve performances from contemporary Deep Learning classifier alone.

BibTeX

@inproceedings{Varma2020Autonomous,
  author = {Varma, Kavita and Nyman, Lisa and Tountas, Konstantinos and Sklivanitis, George and Nayak, Aditya R. and Pados, Dimitris A.},
  booktitle = {Global Oceans 2020: Singapore – U.S. Gulf Coast},
  title = {Autonomous Plankton Classification from Reconstructed Holographic Imagery by L1- PCA-assisted Convolutional Neural Networks},
  year = {2020},
  volume = {},
  number = {},
  pages = {1-6},
  abstract = {Studying and monitoring plankton distribution is vital for global climate and environment protection as they are the most elementary part of oceanic eco-systems. However, the conventional methods and techniques used for understanding the planktons are slow and lacks precision and therefore, in modern day scientific and engineering implementations, Convolutional Neural Networks is extensively used in deep learning and machine learning applications as it outperforms traditional manual approach. Dynamic nature of oceans make it very challenging to monitor these microscopic organisms. Our approach here is to generate a powerful automated plankton recognition system to autonomously identify them and improve the D-CNN for classification of the Plankton holographic imagery curated with the method of Data Conformity Evaluation. The performance of D-CNN classifier is improved by various hyper-parameter tuning, regularization techniques and appending meta-data. Conformity evaluation is based on a matric that's calculated on a continuously refined sequence of calculated L1-norm tensor subspaces of the Plankton images. We note that our classifier performs accurately where our results improve performances from contemporary Deep Learning classifier alone.},
  keywords = {Deep learning;Tensors;Oceans;Training data;Convolutional neural networks;Monitoring;Tuning;Autonomous Classification;D-CNN;Data Conformity;L1-PCA HOLOCAM},
  doi = {10.1109/IEEECONF38699.2020.9389240},
  issn = {0197-7385},
  month = {Oct},
}