Gourishankar Bansode
← Back to Projects

Deep Learning · 2025

Perceptual Loss Engineering & Turbulence-Robust Metric

Research

Systematic study of VGG-based perceptual losses under shifts, noise, and depth-aware atmospheric turbulence, with a new turbulence-robust weighting scheme.

Highlights

  • Compared pixel-wise losses with VGG-16 feature losses and showed 33–92× lower sensitivity to small translations and much higher robustness to additive noise.
  • Integrated a depth-aware atmospheric turbulence simulator to create a dataset of turbulence-degraded images with realistic, depth-dependent distortions.
  • Designed a multi-layer perceptual loss with optimized layer weights that reduced turbulence sensitivity by roughly 24% compared to uniform VGG-layer weighting.