Natural fog and industrial smoke severely degrade optical imaging visibility, undermining real-time applications like autonomous driving and UAV surveillance. This makes haze removal a critical research challenge. This letter proposes PEU-Net (Polarimetric Efficient U-Net), a lightweight Stokes-constrained framework featuring a symmetrical U-Net encoder-decoder with optimized baseline channels, an ECA (Efficient Channel Attention) module for dynamic channel calibration, and a Stokes-based loss function to suppress nonlinear noise. Ablation studies validate 1.3 dB/2.29 dB gains from ECA and Stokes loss, respectively. The framework achieves real-time processing at 33.98 FPS for 1024×1024 inputs and attains 23.04 dB PSNR and 0.821 SSIM on controlled fog chamber datasets, outperforming state-of-the-art methods in accuracy and efficiency. This research provides an efficient solution for real-time polarization-based dehazing tasks and offers novel insights into the design of lightweight physics-data jointly driven models.