Weather Translation via Weather-Cue Transferring

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image translation and restoration tasks have advanced, but multicategory weather translation remains challenging.
• Existing methods struggle with complex semantic structures in weather images and preserving weather-invariant areas.
• A single network capable of handling diverse weather types is needed.

Purpose of the Study:

• To propose a novel weather translation task for transferring image weather types.
• To develop an effective method for multicategory weather translation that preserves image integrity.
• To address the limitations of current approaches in handling complex weather semantics and invariant regions.

Main Methods:

• Developed WeatherGAN, a weather-cue guided multidomain translation approach based on StarGAN v2.
• Redesigned the core generator to adapt weather cues based on the target weather type.
• Introduced a weakly supervised weather segmentation module and a weather clues module for precise area identification.

Main Results:

• WeatherGAN successfully transfers weather cues across multiple categories.
• The approach effectively distinguishes and preserves weather-invariant areas during translation.
• Extensive evaluations demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed WeatherGAN model offers a robust solution for image weather translation.
• The method enhances the ability to edit image weather while maintaining semantic consistency.
• Future work will involve releasing the data and source code for broader research use.