PDDA: Prompt-Driven Domain Adaptation for Real-World Image Dehazing

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Real-world image dehazing is challenging due to domain gaps between synthetic and real data.
• Existing methods struggle with the diversity and complexity of practical environments.
• Bridging the synthetic-to-real domain gap is crucial for robust dehazing.

Purpose of the Study:

• To propose a Prompt-driven Domain Adaptation (PDDA) framework for effective real-world image dehazing.
• To address the scarcity of paired real-world haze images.
• To enhance the cross-domain adaptability and generalization of dehazing models.

Main Methods:

• Introduced a bi-level optimization framework with hyperparameter optimization.
• Developed learnable haze prompts using CLIP latent space for unpaired real haze images.
• Constructed an unsupervised cross-domain loss function integrating prompt learning and bi-level optimization.

Main Results:

• Achieved significant quantitative and qualitative improvements in diverse dehazing scenarios.
• Demonstrated robust performance in real-world daytime conditions.
• Showcased superior cross-domain adaptation capabilities, particularly in nighttime scenarios.

Conclusions:

• The proposed PDDA framework effectively bridges the synthetic-to-real domain gap in image dehazing.
• PDDA exhibits architecture-irrelevant flexibility and domain-agnostic robustness.
• The method offers a promising solution for challenging real-world image restoration tasks.