Classification of marine microalgae using low-resolution Mueller matrix images and convolutional neural network

Area of Science:

• Optics
• Microscopy
• Machine Learning

Background:

• Accurate classification of marine microalgae is crucial for ecological and biotechnological applications.
• Traditional methods often struggle with low-resolution or limited image data.
• Mueller matrix imaging offers rich polarization information potentially useful for microalgae identification.

Purpose of the Study:

• To investigate the effectiveness of convolutional neural networks (CNNs) for classifying marine microalgae using low-resolution Mueller matrix images.
• To compare the classification performance of full Mueller matrix images against standard M11 images at varying resolutions.
• To assess the contribution of polarization information to microalgae classification accuracy.

Main Methods:

• Acquired Mueller matrix images of 12 microalgae species (5 families) using Mueller matrix microscopy at 514 nm.
• Generated seven resolution levels using bicubic interpolation.
• Trained and evaluated CNNs for 12-class (species) and 5-class (family) classification using both full Mueller matrix and M11 images.

Main Results:

• Classification accuracy of Mueller matrix images degraded slowly with decreasing resolution, unlike M11 images which showed sharp declines.
• Mueller matrix images consistently outperformed M11 images across all resolution levels.
• At the lowest resolution, Mueller matrix images provided a 29.89% (species) and 35.83% (family) accuracy improvement over M11 images.

Conclusions:

• Polarization information derived from Mueller matrix images significantly enhances the accuracy of low-resolution marine microalgae classification.
• CNNs are effective tools for analyzing polarization data for microalgae identification.
• The findings highlight the potential of Mueller matrix imaging for robust microalgae classification in challenging conditions.