Digital holographic imaging and classification of microplastics using deep transfer learning.

Area of Science:

• Environmental Science
• Computer Science
• Optical Engineering

Background:

• Microplastic pollution is a growing environmental concern requiring efficient detection methods.
• Current methods for microplastic analysis often struggle with small and imbalanced datasets.
• Digital holographic imaging offers potential for in situ particle analysis but requires advanced processing.

Purpose of the Study:

• To develop and validate a novel deep learning approach for microplastic detection and classification.
• To enhance the performance of digital holographic imaging systems for environmental monitoring.
• To address challenges associated with small and imbalanced datasets in microplastic analysis.

Main Methods:

• An inline digital holographic imaging system was integrated with a lightweight deep learning network (CompNet).
• CompNet features a compression block with concatenated rectified linear unit (CReLU) activation for channel reduction.
• Transfer learning and a class-balanced cross-entropy loss function were employed for training.

Main Results:

• The CompNet system demonstrated significant improvements in feature extraction and generalization.
• Classification accuracy for microplastics was substantially enhanced, effectively mitigating overfitting.
• The approach proved highly suitable for small and imbalanced microplastic datasets.

Conclusions:

• The developed CompNet deep learning method offers a robust solution for in situ microplastic detection and classification.
• This approach shows promise for advancing environmental monitoring technologies.
• The technique effectively overcomes common challenges in analyzing complex environmental samples.