HarmoDehazeDiffusion: Multi-Scale Spatial Fusion Network for Dense Haze Removal

Image dehazing remains a critical yet challenging task in computer vision. While the existing dehazing methods have achieved notable progress, they still struggle to effectively eliminate dense haze in scenes while maintaining consistent illumination between dehazed regions and their surroundings. To address these challenges, we propose a novel image dehazing framework named HarmoDehazeDiffusion. Our model first acquires a depth map of an input image and then integrates these depth cues with hazy images to guide the diffusion denoising process. To simultaneously tackle dense haze removal and illumination consistency, we innovatively incorporate a multi-scale feature extraction strategy, coupled with an adaptive spatial fusion mechanism. Extensive experiments conducted on multiple public benchmarks demonstrate that HarmoDehazeDiffusion achieves a state-of-the-art performance. Quantitative and qualitative evaluations reveal our method’s superior capability to remove haze while preserving natural consistency in the illumination compared to that of the existing approaches, particularly in challenging dense haze scenarios.