MSNet: A Multistage Network for Lightweight Image Dehazing with Content-Guided Attention and Adaptive Encoding

Image dehazing is a critical technique aimed at improving the visual clarity of images. The diverse nature of hazy environments poses significant challenges in developing an efficient and lightweight dehazing model. In this paper, we design a multistage network (MSNet) with content-guided attention and adaptive encoding. The multistage dehazing framework decomposes the complex task of image dehazing into three distinct stages, thereby substantially reducing model complexity. Additionally, we introduce a content-guided attention mechanism that assigns varying weights to different image content elements based on their specific characteristics, thereby improving the efficiency of nonhomogeneous dehazing. Furthermore, we present an adaptive encoder that employs a dual-branch feature extraction structure combined with a gating mechanism, enabling dynamic adjustment of the interactions between the two branches according to the input image. Extensive experimental evaluations on three popular dehazing datasets demonstrate the effectiveness of our proposed MSNet.