Video Dehazing Network Based on Multiscale Attention

Learning-based methods are mostly performed on images. However, less attention has been paid to these methods for video dehazing. This is because of the challenges involved in finding ways to better extract and fuse the spatial and temporal features between consecutive frames, as well as the reconstruction of a more realistic latent frame. This study proposes a multiscale attention video dehazing network (MAVDN) to recover clear dehazed videos. In terms of feature extraction, the proposed method used a feature extractor based on multiscale attention to extracting features at different levels comprehensively. A pixel attention-guided multi-frame fusion module was designed for temporal modeling to aggregate complementary information between adjacent frames. Additionally, a reconstruction module based on cascaded dilated convolutions was designed to better reconstruct the latent frames in the reconstruction stage. Extensive experiments conducted on real-world video dehazing datasets show that the proposed MAVDN achieved superior dehazing performance compared to state-of-the-art methods. Specifically, the PSNR and SSIM of MAVDN results reached
$24.01d\text{B}$
and 0.8832, respectively.