Real-Captured Paired Dataset for Nighttime Flare Removal

Dafeng Zhang

Keywords: Nighttime Flare Removal

Abstract: Flare removal methods eliminate reflective and scattering flares within images and commonly adopt synthetic data for training. However, they fail to achieve robustness for real-world flare-corrupted images as the synthetic data remains gaps with real-world data. In this paper, we propose a real-captured paired dataset named FlareReal600, which contains both real-captured image pairs and pure flare images. Compared with the existing flare removal dataset Flare7k++, our dataset is particularly effective for real-world scenarios as our data contains the faithful mapping between real flare-corrupted images and real flare-free images. Additionally, previous methods either lack sufficient receptive fields or achieve them with huge computational costs, which leads to flares being partly removed or hardly processing high-resolution images. Therefore, we propose a novel flare removal network named \textbf{M}utual re\textbf{C}eption f\textbf{LA}re \textbf{RE}moval \textbf{N}etwork (McLaren), which utilizes convolutions with diverse kernel sizes and fuses them from the perspective of both spatial and channel dimensions to achieve a sufficient receptive field. Furthermore, we employ a re-parameterization mechanism to avoid occupying excessive computational resources. We conduct extensive experiments to demonstrate the functions of our FlareReal600 dataset and our McLaren network.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5563