Single-Image Dehazing via Compositional Adversarial Network.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Image degradation from atmospheric aerosols is a common problem.
• Existing single-image dehazing methods often use separate pipelines for parameter estimation, leading to inefficiencies and errors.

Purpose of the Study:

• To develop a novel generative adversarial network (GAN) for efficient and accurate single-image dehazing.
• To address limitations of existing methods by integrating parameter estimation into an end-to-end pipeline.

Main Methods:

• A novel compositional generator, combining fine-scale and coarse-scale information, is proposed.
• A novel deeply supervised discriminator enforces similarity between generated and clean images at multiple levels.
• The network learns physical parameters directly from data for end-to-end image recovery.

Main Results:

• The proposed GAN method achieves state-of-the-art performance in single-image dehazing.
• The method simultaneously outputs clean images, transmission maps, and air-light values.
• The new HazeCOCO dataset, the largest for single-image dehazing, was created.

Conclusions:

• The novel end-to-end GAN offers a significant advancement in single-image dehazing.
• The integrated approach improves model interpretability and performance.
• The HazeCOCO dataset will facilitate future research in the field.