Rajnandini et al. (2026) Image Restoration via Atmospheric Scattering Models and Deep Learning Techniques

Identification

• Journal: Lecture notes in networks and systems
• Year: 2026
• Date: 2026-01-01
• Authors: Bidisha Rajnandini, Bam Bahadur Sinha
• DOI: 10.1007/978-3-032-10756-5_6

Research Groups

• Computer Science and Engineering Department, National Institute of Technology, Ravangla, South Sikkim, India

Short Summary

This research introduces and evaluates two deep-learning models, DehazeNet and GMAN, for single image haze removal, demonstrating GMAN's superior performance in reconstructing clear images without relying on atmospheric models.

Objective

• To develop and evaluate deep-learning models for single image haze removal, specifically comparing a model-based approach (DehazeNet) with a direct reconstruction approach (GMAN) to overcome limitations of traditional and existing deep learning methods.

Study Configuration

• Spatial Scale: Image level (pixel-wise processing for haze removal).
• Temporal Scale: Instantaneous (processing of a single static image).

Methodology and Data

• Models used:
  • DehazeNet: Convolutional Neural Network (CNN) leveraging the atmospheric scattering model, specialized layers integrating haze-related assumptions and priors, Bilateral Rectified Linear Unit (BReLU) activation function, and Maxout units.
  • GMAN: Encoder-decoder CNN architecture incorporating residual learning at both fine and coarse scales.
• Data sources: Image datasets (specific datasets not mentioned in the provided text).

Main Results

• GMAN significantly outperforms DehazeNet (also referred to as HazeNet) in image restoration tasks.
• Quantitative evaluation showed GMAN achieved a Peak Signal-to-Noise Ratio (PSNR) of 28.59, compared to DehazeNet's 22.94.
• GMAN also demonstrated superior performance in Structural Similarity Index Measure (SSIM) and Universal Image Quality Index (UIQI) metrics.
• GMAN learns directly from data, effectively avoiding common issues of traditional methods such as image darkening or over-sharpening edges.

Contributions

• Introduction of GMAN, a novel deep learning model for single image dehazing that directly reconstructs clean images using an encoder-decoder CNN with residual learning, without explicit reliance on haze-related parameter estimation or atmospheric models.
• Quantitative demonstration of GMAN's superior performance over both traditional methods and model-based deep learning approaches like DehazeNet.
• Development of DehazeNet, a deep learning model that integrates the atmospheric scattering model with novel components like the Bilateral Rectified Linear Unit and Maxout units for enhanced image clarity.

Funding

• Not specified in the provided text.

Citation

@article{Rajnandini2026Image,
  author = {Rajnandini, Bidisha and Sinha, Bam Bahadur},
  title = {Image Restoration via Atmospheric Scattering Models and Deep Learning Techniques},
  journal = {Lecture notes in networks and systems},
  year = {2026},
  doi = {10.1007/978-3-032-10756-5_6},
  url = {https://doi.org/10.1007/978-3-032-10756-5_6}
}