Abstract
Nighttime image flare occurs due to light bending through the lens, creating a high contrast between bright light sources and dark backgrounds, which degrades image quality. To address this, we propose FBNet (Flare Basis Latent Space Transformation Network), a novel network designed for efficient nighttime flare removal. FBNet employs a mapping-based self-attention mechanism to detect and eliminate flare artifacts. Extensive experiments on synthetic and real-world datasets show that FBNet outperforms existing methods, achieving significant improvements in image clarity and quality. Specifically, FBNet is comprehensively evaluated on four datasets, encompassing both synthetic and real-world images, a multi-device dataset featuring 10 consumer-grade cameras, and diverse illumination scenarios including both daytime and nighttime conditions. Our method demonstrates robust generalization and consistently enhances image quality by effectively suppressing flare artifacts, thereby offering a practical and reliable solution for visual enhancement in low-light environments. The code is publicly available at: https://github.com/Leo-student/Flare7K
Funding Information
  • China NSF (62427801)
  • National Key Research and Development Program of China (2023YFB4502802)
  • China NSF (62031009)
  • Fudan-ZTE Joint Lab
  • Alibaba Research Fellow (ARF) Program

This publication has 53 references indexed in Scilit: