Visibility Estimation Based on Weakly Supervised Learning under Discrete Label Distribution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Estimating visibility in fog is crucial for safe navigation, especially for autonomous systems.
• Existing methods often struggle with the complex and uneven distribution of fog.
• Accurate visibility estimation requires understanding regional variations within fogged images.

Purpose of the Study:

• To propose an end-to-end neural network model for accurate visibility estimation in fog images.
• To leverage the characteristic of uneven fog distribution for improved performance.
• To develop a robust method with low annotation requirements for practical applications.

Main Methods:

• Transformed single labels into discrete label distributions for regional analysis.
• Introduced discrete label distribution learning within classification networks.
• Employed a bilinear attention pooling module to identify the farthest visible fog region.
• Utilized an attention-based branch and cascaded feature fusion with a base branch.

Main Results:

• The proposed model effectively estimates visibility by utilizing uneven fog distribution.
• Demonstrated effectiveness on both real highway and synthetic road datasets.
• Achieved low annotation requirements, indicating practical feasibility.
• Showcased good robustness and a broad application space for the visibility estimation method.

Conclusions:

• The developed neural network model provides an effective solution for visibility estimation in fog.
• The discrete label distribution learning and attention mechanisms enhance the model's ability to interpret complex fog patterns.
• The method's robustness and low annotation needs make it suitable for real-world deployment in various scenarios.