Pixel-Wise Wasserstein Autoencoder for Highly Generative Dehazing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image dehazing is crucial for improving visual quality in adverse atmospheric conditions.
• Existing generative adversarial network (GAN)-based methods often lack style diversity and control.

Purpose of the Study:

• To develop a highly generative and accurate image dehazing method.
• To enable diverse style transfer for dehazed images.
• To enhance image quality across various challenging scenarios.

Main Methods:

• Utilized pixel-wise Wasserstein autoencoders for generative dehazing.
• Implemented a 2D latent tensor representation for precise pixel matching.
• Introduced an advanced feature fusion technique for richer latent space information.
• Developed a mapping function for style transfer in latent spaces.

Main Results:

• Achieved significant improvements in dehazing accuracy compared to state-of-the-art methods.
• Generated highly diverse haze-free images with varied tones, illuminations, and moods.
• Demonstrated superior quantitative performance on synthetic and real-world datasets.

Conclusions:

• The proposed Wasserstein autoencoder-based method offers a powerful and flexible approach to image dehazing.
• The technique enables diverse stylistic outputs and broad applicability, including low-light, nighttime, and underwater image enhancement.