High Dynamic Range Image Recovery by Use of Lens Flare Events Detection Algorithm

Event cameras are novel vision sensors that asynchronously output per-pixel events by measuring brightness changes. Event cameras have advantages such as high speed and high dynamic range compared to conventional cameras. To leverage the advantages and apply prior algorithms, event-based image reconstruction has been developed. With the development of neural networks, state-of-the-art reconstruction methods are introduced. However, high dynamic range reconstructions still suffer from lens flare-based artefacts, which makes the intensity estimation incorrect. To address this problem, this work presents a computational approach to detect ill-posed events caused by lens flare. Under the examination that lens flare elements of event cameras mainly consist of glow and starburst, we derive two bivariate Gaussian distributions from targets in the compressed stream of events. By operating convolution, the detector reduces the sparsity of events, which makes the surface fitting more precise. We show that the proposed method effectively eliminates dark stain in high dynamic range reconstruction, while preserving detail on the other region at the same time.