<p>Image dehazing enhances visibility and clarity by removing haze, which is crucial for computer vision activities such as object detection, tracking, and recognition, as haze significantly degrades image quality and hinders accurate interpretation. Conventional dehazing techniques rely on hand-crafted priors and assumptions which can falter under a variety of real-world circumstances and produce inconsistent outcomes with limited effectiveness in scenarios involving complex or dense haze. Recent deep learning approaches, though effective, either focus solely on spatial or frequency features, limiting their generalization and visual realism. In this paper, we introduced a novel image dehazing architecture, MHAze-Net, that integrates an Adaptive Multi-Head Residual UNet as the generator and a Multi-Scale Attention Fusion Network as the discriminator. The generator utilizes multi-head attention mechanisms to improve feature extraction, while the discriminator incorporates both high and low-frequency information as priors and constraints. It employs spatial attention-guided mechanisms to enhance its capacity to differentiate between actual and generated dehazed images. Comprehensive experiments on both real-world and synthetic datasets highlight the effectiveness of our model, which consistently achieves superior PSNR and SSIM values when compared to existing state-of-the-art approaches. Additionally, we provide a detailed analysis of computational efficiency, evaluating runtime, parameter count, and FLOPs to demonstrate the model’s lightweight design. The practical applicability of our framework is further validated through successful license plate detection in hazy traffic scenes, underscoring its relevance in real-world scenarios.</p>