Real-Time Classification of Microplastic Particles Using an Efficient Encoder for Holographic Images

Microplastic (MP) pollution in marine environments poses a significant ecological threat, impacting biodiversity and human health through bioaccumulation and ecosystem disruption. Traditional monitoring methods are labour-intensive, geographically constrained, and lack scalability for global assessments. The task remains challenging due to complex morphologies, dust contamination, and class imbalance in real-world samples. We present a deep learning based classification of Microplastic particles using raw holographic imaging without numerical reconstruction in water samples. We propose a robust framework that is built on the base EfficientNet-B2 encoder enhanced with a custom classification head. The proposed model is trained on a 9-class holographic dataset utilizing advanced data augmentation techniques, label smoothing to ensure generalization, and prevent overfitting. Evaluation is carried out using stratified 5-fold cross-validation, yielding a mean accuracy of 0.9944. Furthermore, it enables real-time inference at 8.3 ms per image. The results suggest that the proposed framework offers an efficient and reproducible solution for automated MP analysis, with potential applicability in environmental monitoring and laboratory-scale screening workflows.