Toward Blind Flare Removal Using Knowledge-Driven Flare-Level Estimator.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Lens flare introduces artifacts in images, complicating image analysis and restoration.
• Existing deep learning methods for flare removal are limited by scarce real-world training data.
• Synthesized training data often fails to capture the complexity of real-world flare phenomena.

Purpose of the Study:

• To develop a novel, knowledge-driven method for effective lens flare removal.
• To address the limitations of current deep learning approaches due to insufficient real image pairs.
• To improve the adaptive performance of flare removal networks during training and testing.

Main Methods:

• A simple yet effective flare-level estimator predicts image corruption levels.
• A flare-level modulator integrates estimated flare information into neural networks.
• A flare-aware block is designed for enhanced flare recognition and reconstruction.
• A new real-world dataset, WiderFlare, was collected for benchmarking.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art techniques.
• Quantitative and qualitative evaluations across three benchmark datasets confirm effectiveness.
• The flare-level estimator adaptively aids network performance in both training and testing.

Conclusions:

• The knowledge-driven approach offers a significant advancement in blind flare removal.
• The developed flare-level estimator and flare-aware block effectively improve restoration accuracy.
• The WiderFlare dataset provides a valuable resource for future research in lens flare removal.