Dual-Scale Single Image Dehazing via Neural Augmentation.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Model-based dehazing offers sharp details but low synthetic image quality.
• Data-driven dehazing yields high synthetic quality but struggles with real-world haze and contrast.
• Existing methods present trade-offs between real-world and synthetic image restoration.

Purpose of the Study:

• To develop a novel single image dehazing algorithm.
• To combine the strengths of model-based and data-driven approaches.
• To improve haze removal for both real-world and synthetic images.

Main Methods:

• Estimating transmission map and atmospheric light using model-based techniques.
• Refining these parameters with dual-scale Generative Adversarial Networks (GANs).
• Restoring haze-free images via Koschmieder's law.

Main Results:

• The proposed algorithm demonstrates effective haze removal across diverse image types.
• Achieved fast convergence due to neural augmentation.
• Outperforms existing methods by balancing real-world and synthetic image restoration.

Conclusions:

• The hybrid approach successfully integrates model-based and data-driven strengths.
• The novel algorithm provides superior dehazing performance.
• This method offers a robust solution for single image dehazing challenges.