Adaptive-frequency-guided atmospheric turbulence mitigation for long-range optical video imaging

Atmospheric turbulence degrades long-range optical video imaging through time-varying distortion, spatially varying blur, and temporal instability. We propose AUTO, an adaptive-frequency-guided framework for turbulence video restoration. AUTO enhances turbulence-sensitive structural features, aligns distorted frames with guided deformable alignment, aggregates temporal information bidirectionally, and separates geometric stabilization from detail recovery. The weak, medium, and strong regimes in ATSyn_dynamic follow the original benchmark protocol and are defined by average pixel displacement, blur-kernel basis size, and D / r 0 rather than C n 2 alone. On ATSyn_dynamic, AUTO achieves the best overall PSNR of 29.17 dB and competitive SSIM. Real text sequences and self-collected, mostly static near-ground videos are further used to assess text readability and qualitative/no-reference real-scene restoration.