SemiDDM-weather: A semi-supervised learning framework for all-in-one adverse weather removal

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Adverse weather conditions significantly degrade visual quality in images and videos.
• Current weather removal techniques often require extensive labeled data and are specific to certain weather types.
• Developing a versatile and data-efficient solution for all-in-one adverse weather removal is crucial.

Purpose of the Study:

• To present a pioneering semi-supervised framework for all-in-one adverse weather removal.
• To overcome the limitations of data dependency and weather-type specificity in existing methods.
• To enhance the efficiency and effectiveness of adverse weather removal using limited labeled data.

Main Methods:

• A teacher-student network architecture is employed with a Denoising Diffusion Model (DDM) as the backbone (SemiDDM-Weather).
• The Wavelet Diffusion Model (Wavelet-Diffusion) is adapted with customized inputs and loss functions for efficient many-to-one mapping.
• Quality assessment and content consistency constraints are introduced to refine pseudo-labels for robust semi-supervised learning.

Main Results:

• SemiDDM-Weather demonstrates high visual quality and superior performance in adverse weather removal across synthetic and real-world datasets.
• The framework effectively handles diverse weather conditions, outperforming fully supervised methods.
• The semi-supervised approach significantly reduces the need for extensive labeled data.

Conclusions:

• SemiDDM-Weather offers an effective and efficient solution for all-in-one adverse weather removal.
• The proposed quality assessment and content consistency mechanisms improve semi-supervised learning robustness.
• This framework advances the field by enabling high-quality image restoration under adverse conditions with reduced data requirements.