Rethinking mixture of rain removal via depth-guided adversarial learning

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Rainy weather severely degrades outdoor image visibility due to artifacts like raindrops, streaks, and haze.
• Existing image deraining methods struggle with the mixture of rain (MOR) due to limited artifact handling.
• Degradation varies with distance, complicating effective rain removal.

Purpose of the Study:

• To propose an effective image deraining paradigm for Mixture of Rain (MOR) removal.
• To develop a model that leverages scene depth information for enhanced deraining.
• To introduce a novel normalization technique to boost deraining performance.

Main Methods:

• A joint learning paradigm, DEMore-Net, integrating depth estimation and MOR removal.
• Utilizing scene depth as guidance for differentiating and removing various rain artifacts.
• Introducing a Hybrid Normalization Block (HNB) to improve deraining efficacy.

Main Results:

• DEMore-Net demonstrates superior performance in removing mixtures of rain compared to existing methods.
• Depth information effectively guides the removal of diverse rain artifacts.
• Experiments on synthetic and real-world data validate the model's effectiveness.

Conclusions:

• DEMore-Net offers a significant advancement in tackling complex image degradation caused by rain.
• Integrating depth estimation is a promising direction for robust image deraining.
• The proposed Hybrid Normalization Block further enhances the model's deraining capabilities.