An Image Reconstruction Method on Brightness Constancy for Event Data

Compared to conventional frame-sampling sensors, Dynamic Vision Sensors can rapidly and asynchronously capture brightness changes in a scene, outputting discrete and sparse events. Despite their innovative operational paradigm conferring distinctive benefits, they also reveal problems such as the non-evident graphical representation of event streams, the absence of textural information within the documented data, and the need for developing compatible algorithms. To address these issues, we propose an image reconstruction neural network that considers the brightness consistency assumption. The method integrates the working principle of event cameras into a deep learning network, optimizing the brightness estimation process under the constraint of brightness consistency principle, combined with event-based optical flow estimation. By training the network model, we achieve the reconstruction process from pure event streams to high-resolution intensity images. Finally, we compare our method with MR, E2VID and FireNet algorithms. On the calibration dataset, our advanced method demonstrates the lowest Mean Squared Error of 12.55%, and the Structural Similarity that surpasses the remaining three methods at 91.76%, substantiating the superior efficacy of the proposed algorithm.