<p>The growing interest in image haze removal has driven significant advancements in deep learning-based algorithms, exhibiting promising performance in recent studies. However, due to the uneven and complex distribution of haze in real-world scenarios, existing single-scale models are constrained by their restricted receptive fields, making it challenging to effectively address intricate hazy environments. Besides, conventional down-sampling operations may easily suffer from loss of details and color distortion when reducing the feature resolution, hindering the realization of better haze removal performance. To this end, we propose a Multi-scale Wavelet-guided Attention Network (MWA-Net) for haze removal. For one thing, our MWA-Net relies on an efficient multi-stage architecture that progressively enlarges receptive fields from multi-scale degraded hazy inputs, effectively capturing global content information for better haze removal in complex scenarios. Meanwhile, we incorporate a novel mixed attention mechanism that includes Channel Attention, Pixel Attention, and Supplementary Attention to further boost the feature learning for alleviating uneven haze distribution. For another, we leverage wavelet decomposition to extract low-frequency and high-frequency components from the feature layer for feature down-sampling and up-sampling, which enables the model to minimize texture detail loss and ensure high-quality feature reconstruction. Furthermore, an Attentive Fusion Scheme is designed to effectively integrate multi-scale features while overcoming the semantic and scale inconsistency issues among diverse inputs. Extensive experiments prove that our method MWA-Net can outperform against other compared state-of-the-art methods on several benchmarks.</p>