Video reconstruction through dynamic scattering media based on physics-informed spatio-temporal transformer

Dynamic scattering media video reconstruction represents one of the most challenging problems in computational imaging, with applications ranging from autonomous driving under adverse weather conditions to biomedical imaging through turbid tissues. Traditional approaches are often limited in their ability to capture the complex temporal dynamics and physics-based constraints inherent in scattering phenomena, which can result in reduced reconstruction quality and restricted applicability. In this work, we present PISTA (Physics-Informed Spatio-Temporal Transformer Architecture), a deep learning framework that integrates physical principles with advanced attention mechanisms for improved video reconstruction through dynamic scattering media. Our approach addresses three fundamental limitations of existing methods: incomplete modeling of temporal correlations in scattering dynamics, insufficient incorporation of physics-based constraints into neural network architectures, and limited parameter estimation for time-varying scattering properties. PISTA employs a physics-informed encoder that enforces energy conservation, temporal causality, and reciprocity, coupled with a spatio-temporal attention module that captures long-range dependencies across both spatial and temporal dimensions. Furthermore, we introduce a parameter estimation network that adaptively learns scattering coefficients, enabling reconstruction that is responsive to dynamic medium variations. We validate the proposed framework on both physics-based synthetic data and the OTIS real turbulence dataset, demonstrating notable improvements over conventional CNN- and transformer-based baselines. Overall, PISTA provides a physically consistent and data-efficient solution for video reconstruction in dynamic scattering environments.