When low-light meets flares: Towards Synchronous Flare Removal and Brightness Enhancement

Area of Science:

• Computer Vision
• Image Processing

Background:

• Low-light image enhancement (LLIE) aims to improve visibility but struggles with flares in real-world images.
• Existing methods face challenges like uneven lighting, noise amplification, and loss of detail in dark regions.

Purpose of the Study:

• Propose a novel task: Flared Low-Light Image Enhancement (FLLIE).
• Develop an effective deep learning network for FLLIE.

Main Methods:

• Synthesized flared low-light datasets for training and inference.
• Developed a Fourier transform-based deep network: Synchronous Flare Removal and Brightness Enhancement (SFRBE).
• Introduced a Residual Directional Fourier Block (RDFB) for frequency-domain feature extraction.

Main Results:

• Demonstrated the effectiveness of the SFRBE network on synthesized and real-world flared low-light images.
• The proposed method successfully tackles challenges of flares, noise, and detail preservation.

Conclusions:

• The novel FLLIE task and SFRBE network offer a significant advancement in handling complex low-light conditions.
• SFRBE provides a robust solution for enhancing images with both low illumination and light flares.