Enhanced convolutional neural network for plankton identification and enumeration.

Area of Science:

• Marine biology
• Image processing
• Computational science

Background:

• Plankton imaging systems are increasingly used in marine science.
• Processing large plankton image datasets is challenging due to variations in image quality and content.
• Automated recognition and enumeration are crucial for efficient data analysis.

Purpose of the Study:

• To develop an automated system for plankton image recognition and enumeration.
• To evaluate the performance of different Convolutional Neural Network (CNN) structures for plankton classification.
• To compare the effectiveness of a combined CNN-Support Vector Machine (SVM) classifier against CNN alone.

Main Methods:

• An adaptive thresholding approach was used to extract Regions of Interest (ROIs).
• Background noise suppression and feature enhancement were applied to ROIs.
• A pre-trained classifier, combining a CNN for feature description and an SVM for classification accuracy, was employed.
• ResNet50 was identified as the best-performing CNN model.

Main Results:

• The CNN-SVM combination improved classification accuracy by 7.13% and recall by 6.41% compared to CNN alone.
• The ResNet50 model achieved the highest accuracy (94.52%) and recall (94.13%).
• The developed algorithm demonstrated general applicability across different imaging systems.

Conclusions:

• Deep learning techniques significantly enhance plankton image recognition and enumeration.
• The study provides valuable insights into selecting appropriate CNN models for plankton analysis.
• The proposed automated system offers a robust solution for processing large-scale plankton image data in marine science.