Toward accurate single image sand dust removal by utilizing uncertainty-aware neural network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning shows promise in image de-dusting but struggles with heterogeneous environmental uncertainty.
• Existing methods often fail to adequately address the complex degradations caused by sand and dust.

Purpose of the Study:

• To develop a novel deep learning framework, the Hierarchical Interactive Uncertainty-aware Network (HIUNet), for robust single image sand and dust removal.
• To effectively model and mitigate the uncertainty inherent in dusty environments for improved image restoration.

Main Methods:

• Utilized Bayesian neural networks for robust shallow feature extraction.
• Employed pre-trained encoders and lightweight decoders for initial image reconstitution.
• Implemented a feature frequency selection mechanism to identify and retain valuable features while suppressing noise.
• Integrated a feature enhancement module for refining the preliminary restoration.

Main Results:

• HIUNet demonstrated superior performance in reconstructing high-quality clean images from degraded inputs.
• Experiments on the Sand11K dataset validated the method's effectiveness across various degradation levels.
• The framework successfully modeled uncertainty and selected salient features for restoration.

Conclusions:

• HIUNet offers an effective solution for single image sand and dust removal by addressing environmental uncertainty.
• The combination of Bayesian networks and feature frequency selection is key to high-quality image reconstruction.
• Future work will focus on extending the framework to extreme sand scenarios.