Encoding Holographic Data Into Synthetic Video Streams for Enhanced Microplastic Detection

This study presents a novel deep learning pipeline for the detection and classification of microplastics using digital holography, with a focus on synthetic microfibers released during laundry. We introduce HMPD 2.0, an enhanced version of the Holography MicroPlastic Dataset featuring 24 amplitude and phase channels per sample, obtained through varying spatial filtering, numerical aberration correction, and propagation. To reduce input dimensionality, we propose a pseudo-RGB compression technique that groups grayscale channels into synthetic frames, which are then interpreted as a video sequence. This allows the use of transformer-based video architectures, particularly TimeSformer, for spatiotemporal modeling. Experimental results demonstrate that TimeSformer achieves a classification accuracy of up to 97.91%, with compressed 8-frame inputs maintaining high performance while significantly reducing inference time (from 42 ms to 16 ms per sample). These findings validate the effectiveness and efficiency of our approach, which supports real-time deployment on edge devices.