Lightweight Multi-scale Attentional Network for Single Image Dehazing

Abstract

Neural network-based image dehazing algorithms have obtained extensive research results. However, haze retention is still present in the results for complex outdoor images. Designing complicated networks may be improved, yet they consist of a tremendous number of parameters, which causes slow inference and requires a more demanding computer environment for execution. The network globally adopts a U-Net-like structure and employs two improved modules ARM and SAM based on the attention mechanism to implement multi-scale feature extraction and fusion. Our proposed network combines the high and low level semantics in diverse scale feature images while raising the focus on diverse haze concentration regions. Experimental results reveal that our network can get improved visual results in RESIDE-SOTS (Outdoor) and NTIRE2021 NH-HAZE Image Dehazing Challenge datasets in comparison with other SOTA methods and attains higher Peak Signal to Noise Ratio and Structural Similarity Index.