Abstract: An event camera is a neuromorphic asynchronous sensor that can output event data with high
dynamic range and no motion blur at microsecond-level temporal resolution, directly capturing change
information in a scene. Deep learning-based image reconstruction from event streams shows promise for
various vision tasks but often suffers from high complexity, high computational cost, or poor
reconstruction quality in overly simplified models. To address these issues, this paper proposes an
efficient event reconstruction network combining a recurrent convolutional architecture with an efficient
channel attention mechanism. Experimental results show that, compared with existing reconstruction
methods, the proposed reconstruction network demonstrates improved performance on different types of
standard event camera datasets, achieving mean squared error, structural similarity index, and perceptual
similarity loss scores of 0.061, 0.60, and 0.38, respectively, while ensuring lightweight and efficient
operation.

Key words: event camera, image reconstruction, neural network, deep learning