End-to-end deep unfolding network for DoFP polarization image reconstruction.

Snapshot polarimetric imaging systems using division of focal plane (DoFP) sensors efficiently capture scene polarization information, but the spatial multiplexing of micro-polarizer arrays leads to a loss of spatial resolution. Current learning-based polarization image reconstruction methods, although achieving excellent performance by directly learning the mapping from low resolution to high resolution, often neglect the inherent physical mechanisms of polarization image reconstruction. This paper proposes a deep unfolding network for polarization image reconstruction called DUPIR, which jointly reconstructs full-resolution intensity images for the four linear polarization orientations I0, I45, I90, and I135, along with their corresponding polarization parameters Stokes S0, the degree of linear polarization DoLP, and the angle of polarization AoP, in an end-to-end manner. By integrating physical model priors into a trainable architecture, DUPIR bridges the gap between model-based and learning-based methods. To mitigate the lack of high-quality training data, a polarization dataset comprising 184 sample pairs from various object categories was established. Extensive experiments on both public and our collected datasets demonstrate that DUPIR achieves state-of-the-art reconstruction accuracy while maintaining real-time inference capability.