Learning-based polarization retrieval from intensity speckle of dense scattering media

Polarization information is vital for recovering the fine details and structure of the object in complex scattering environments. However, existing methods for acquiring polarization information typically rely on specialized imaging systems, limiting their broader applications. To simplify the imaging system and reduce computational complexity, we introduce a deep neural network-based approach capable of retrieving the polarization information directly from a single-intensity image captured by a conventional camera. The network integrates transformer and U-Net architectures, incorporating multi-level feature extraction, skip connections, and multi-head self-attention mechanism (MHSA), which facilitate information fusion and allow precise recovery of polarization information from intensity images in highly scattering environments. The proposed method effectively extracts the degree of polarization (DoP) and angle of polarization (AoP) from a single speckle intensity image without requiring a custom system. It achieves 98% accuracy in cosine similarity, demonstrating robust performance in interpreting polarization information in dense scattering media, offering what we believe is a novel solution for imaging through scattering environments.