<p>Image dehazing is a critical task in the field of computer vision, aiming to restore clear scenes from images affected by haze. Although methods based on Convolutional Neural Networks (CNN) and Transformers have made significant progress in this task, existing approaches still exhibit limitations. However, CNNs are limited in capturing global dependencies due to restricted receptive fields, while the high computational cost of Transformers hinders their deployment in real-world applications. Moreover, traditional encoder-decoder architectures often experience information loss and semantic inconsistency during cross-scale feature transmission caused by resolution disparities. To address these challenges, this paper proposes a novel dehazing network, GL-MambaNet, which achieves efficient global-local modeling through a dual-path collaborative mechanism and an adaptive feature fusion strategy. Specifically, we design the Attentive Enhanced Mamba Block (AEMB). To prevent the network from overemphasizing global information while balancing local and global features, we introduce a dual-path structure combined with RepVGG. To tackle the issue of feature asymmetry in encoder-decoder architectures, we propose the Adaptive Depthwise Feature Fusion (ADFF) module. This module dynamically adjusts the weight distribution of cross-scale features using learnable depthwise convolution kernels, effectively eliminating interference from redundant information while preserving multi-scale semantic integrity. Extensive experiments on both synthetic and real-world datasets demonstrate that, compared to state-of-the-art methods, our approach not only achieves superior dehazing performance but also maintains low computational resource consumption, striking an excellent balance between performance and efficiency.</p>