LUFormer : A luminance-informed localized transformer with frequency augmentation for nighttime flare removal

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image flare, caused by light scattering and reflection in night scenes, severely degrades visual quality.
• Existing flare removal techniques often fail to integrate all relevant information, limiting their effectiveness.

Purpose of the Study:

• To propose LUFormer, a novel network for robust image flare removal.
• To address limitations of current methods by comprehensively integrating luminance and frequency information.

Main Methods:

• Developed LUFormer, a luminance-informed Transformer network with localized frequency augmentation.
• Introduced a luminance-guided branch (LGB) for global brightness semantic priors.
• Utilized a dual domain hybrid attention (DDHA) unit for spatial and frequency domain feature enhancement.

Main Results:

• LUFormer demonstrated superior performance over state-of-the-art methods on real-world benchmarks.
• The LGB effectively captures luminance distribution disruptions caused by flare.
• The DDHA improved deep flare representation by broadening receptive fields and amplifying low-frequency components.

Conclusions:

• LUFormer offers a significant advancement in image flare removal technology.
• The proposed method provides a robust solution for enhancing image quality in challenging night conditions.
• The integration of global luminance guidance and localized multi-domain attention is key to the method's success.