CVID

CVID performs image deraining using Conditional Variational Auto-Encoders (CVAEs) to generate probabilistic, diverse reconstructions that improve visual clarity in images affected by rain.

Key Features:

• Conditional Variational Auto-Encoders (CVAEs): Employs CVAEs for probabilistic inference and to produce diverse derained predictions.
• Spatial Density Estimation (SDE) module: Estimates a rain density map per image to enable spatially adaptive deraining.
• Channel-Wise (CW) deraining: Processes each color channel individually to account for channel-specific rain density variations.
• Diversity of predictions: Produces multiple plausible reconstructions to address variability in rain intensity across space and color channels.
• Empirical performance: Reports superior deraining performance compared to deterministic image deraining methods on benchmark datasets.
• Ablation validation: Uses ablation studies to confirm the contributions of the SDE module and CW scheme.

Scientific Applications:

• Remote sensing: Enhances clarity of satellite and aerial imagery captured under rainy conditions to improve downstream analysis.
• Autonomous driving: Improves visibility in vehicle-captured images to support perception and decision-making in rainy environments.
• Surveillance systems: Restores visual quality of security footage affected by rain for more reliable monitoring and analysis.

Methodology:

Integrates probabilistic inference via CVAEs, uses an SDE module to estimate per-image rain density maps, applies a channel-wise deraining scheme processing individual color channels, and evaluates results on synthesized and real-world datasets with ablation studies.