Snapshot video through dynamic scattering medium based on deep learning.

Abstract

We present an end-to-end deep learning model designed to reconstruct up to eight frames from a single snapshot of a dynamic object passing through an unknown, time-varying scattering medium. Our approach integrates a coded aperture compressive temporal imaging system with a specially designed transformer-based convolutional neural network (CNN), optimized for effective demultiplexing and reconstruction. Both simulation and experimental results demonstrate a successful compression ratio of up to 8X, while maintaining high reconstruction quality. Furthermore, ablation studies reveal that our dual-input CNN model, which utilizes both speckle patterns and their autocorrelations, significantly improves reconstruction accuracy.