Image Lens Flare Removal Using Adversarial Curve Learning.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Heterogeneous flare artifacts degrade image quality and impact computer vision tasks.
• Synthesizing training data for flare removal is challenging due to limitations in current methods, such as ignoring automatic exposure and tone mapping in the image signal processing pipeline (ISP).
• Existing light source recovery methods struggle with multiple light sources of varying characteristics.

Purpose of the Study:

• To improve lens flare removal performance by revisiting the ISP and enhancing the synthesis pipeline.
• To develop a more reliable strategy for recovering multiple light sources.
• To enhance the generalization capability of deep flare removal models across different devices.

Main Methods:

• Revisiting the ISP and remodeling automatic exposure principles within the synthesis pipeline.
• Implementing a new light source recovery strategy using convex averaging based on illuminance levels.
• Employing an adversarial curve learning (ACL) paradigm for improved generalization across diverse camera ISPs.

Main Results:

• The proposed pipeline generates more realistic imaging by discriminating local and global illumination, preventing shifts and over-saturation.
• The ACL paradigm enhances model generalization, overcoming limitations of device-specific ISPs.
• The new light source recovery strategy effectively handles multiple light sources without hard thresholds.

Conclusions:

• The developed solution significantly improves lens flare removal performance.
• The approach advances the field towards handling more general imaging situations.
• A new dataset for testing flare removal generalization across various consumer electronics is introduced.