LGL: Local guide local network for non-homogeneous image dehazing

Zhongze Wang, Haitao Zhao, Jingchao Peng, Kaijie Zhao, Lujian Yao

Abstract: Non-homogeneous image dehazing (NHID) presents a significant challenge due to its uneven haze distribution. Information on areas with different densities contributes to the dehazing of other areas, necessitating long-range modeling capabilities in dehazing methodologies. Leveraging the global self-attention, Transformer architectures offer promise for NHID. However, pixel-level token embedding and the quadratic complexity of self-attention pose hurdles for effectively modeling non-homogeneous haze images. In this paper, we propose a novel method tailored for NHID, named Local Guide Local Network (LGL), which selects information with low similarity within each local window to guide the dehazing of other local windows. Specifically, we propose an Orthogonal Token Selector, optimized geometrically on the Stiefel manifold, to select unique tokens with top-k attention values. Subsequently, a Global Sparse Attention is performed among local windows to extract global features and exchange features of local areas based on the selected tokens. Finally, a Local Guided Feedforward Network is proposed to achieve region-specific dehazing. Extensive experimental evaluations validate the efficacy of our LGL for NHID, demonstrating superior performance over existing methods.

External IDs:dblp:journals/ijon/WangZPZY25