Computational Polarimetric Holography for Efficient Microplastic Classification via a Lightweight Wavelet-Enhanced Vision Transformer

Efficient in-situ detection of marine microplastic pollution remains a critical challenge in environmental monitoring. While most existing detection technologies are lab-based and highly accurate, they are also time-consuming, costly, and labor-intensive. In contrast, portable aquatic microplastic detection systems can significantly reduce these overheads. Holographic polarimetry is a promising method with the potential for in-situ detection. However, current research in this area has largely focused on the laboratory-based analysis of microplastic features or utilized networks too computationally intensive for deployment on edge devices. This study aims to enable high-throughput, in-situ detection for multiple classes of microplastics. To this end, we built a portable holographic polarimetric microplastic imager (HPM imager) and developed WMViT3, a lightweight feature extraction model for microplastic classification. Unlike traditional methods, our approach bypasses complex reconstruction by directly decoding the raw polarimetric interference fringes. The model integrates a wavelet transform specifically designed to capture the high-frequency oscillatory features of holographic fringes, effectively extracting unique "optical fingerprints" of different materials. It achieves a classification accuracy of 97.97% on our standardized HPM-500 dataset while reducing computational load by 53.2%. Crucially, this efficiency enables real-time inference on embedded devices, providing a viable technical solution for portable, real-time environmental monitoring systems.