Rich Teacher Features for Efficient Single-Image Haze Removal

Lightweight image dehazing using feature distillation and affinity modules

Rich Teacher Features for Efficient Single-Image Haze Removal

Single-image haze removal is a challenging computer vision task that significantly impacts image quality and downstream applications. This project introduces a novel lightweight framework that leverages heterogeneous knowledge distillation from super-resolution models to achieve efficient and effective image dehazing.

Key Innovation: Feature Affinity Module

Our approach exploits rich “dark-knowledge” information from a lightweight pre-trained super-resolution teacher model through a specially designed Feature Affinity Module that maximizes the flow of semantic features to the student dehazing network.

Technical Approach

  1. Heterogeneous Knowledge Distillation: Transfer knowledge from super-resolution domain to dehazing
  2. Feature Affinity Module: Designed to maximize rich feature semantic flow
  3. Lightweight Architecture: Significantly reduced model complexity while maintaining performance
  4. Plug-and-Play Design: Can be integrated with existing baseline models

Results & Performance

  • PSNR Improvement: Up to 15% gains in Peak Signal-to-Noise Ratio
  • Model Efficiency: ~20× reduction in model size compared to traditional approaches
  • Dataset Validation: Comprehensive evaluation on RESIDE-Standard dataset
  • Robustness: Demonstrated effectiveness on both synthetic and real-world domains

Architecture Highlights

The framework employs a dual-path architecture with:

  • Clear Feature Maps (F.M): Processing clean image features
  • Hazy Feature Maps (F.M): Processing degraded input features
  • Pooling and Flattening: Efficient feature dimensionality reduction
  • L2 Loss Optimization: Robust training objective for feature alignment

Applications

This lightweight dehazing solution is particularly suitable for:

  • Edge Computing: On-device image enhancement
  • Real-time Processing: Low-latency applications
  • Resource-Constrained Environments: Mobile and embedded systems
  • Surveillance Systems: Improving visibility in adverse weather conditions

GitHub Repository: anushrisuresh/Dehaze

Technologies: PyTorch, Computer Vision, Knowledge Distillation, Image Processing, Deep Learning