A Novel Attention-Based Network for Single Image Dehazing

Weihao Gao,Yongjun Zhang,Huachun Jian

DOI: https://doi.org/10.1007/s00371-023-03129-w

2024-01-01

Abstract:As one typical severe weather, haze can exert a bad influence on vision tasks. Therefore, research on the image dehazing task is of great importance. Recent years have witnessed the remarkable success of CNN-based dehazing algorithms, showcasing their powerful capabilities. Nevertheless, despite this progress, these approaches still exhibit limitations in applications, and there remains considerable untapped potential for further refinements. To address this challenge, we develop a novel CNN-based network dubbed (MSSDN), which demonstrates superior dehazing performance and highlights its potential for real-world use cases. Specifically, we design a multi-spectral attention module (MSAM), which shows its superiority in collecting information in the channel dimension. And based on it, a haze capture module (HCM) is designed to eliminate hazy information. Meanwhile, the designed background feature capture module (BFCM) enlarges the receptive field of the proposed network to capture more useful information. Finally, the designed cross stage interaction module (CSIM) promotes the process of information flow, and the contrastive learning is adopted to decouple the haze component from the background part. The evaluation results demonstrate the superiority of the proposed algorithm, and our MSSDN outperforms the SOTA dehazing methods.