【摘要】 图像去雾一直是计算机视觉领域的一个具有挑战性的任务。近年来，随着深度学习的发展，基于卷积神经网络（CNN）的去雾网络得到了广泛应用，现有的基于深度学习的方法在同时有效捕获去雾所需的全局上下文信息与精细结构信息方面仍存在难以兼顾去雾效果与模型效率的局限性问题。提出一种基于CNN的多注意力融合新型图像去雾方法，通过引入空间-通道协同注意力（SCSA）模块，利用门控机制自适应地重校准特征，同时挖掘空间与通道维度之间的依赖关系，从而使模型能够更加聚焦于信息量更高的区域与特征；设计多尺度特征融合（MSFF）模块，通过不同感受野的卷积核提取并聚合特征，使网络能够更好地处理不同尺度的雾化模式。这些模块被嵌入到由多个残差组与跨组特征融合策略构成的深层网络中，以提升去雾效果。在RESIDE等基准数据集上的实验结果表明，所提方法在室内数据集上的峰值信噪比指标达到38.86，在结构相似性指数达到0.9912的同时参数量仅为4.99×10~6，在视觉上效果也能使色彩饱和度与真实值最为接近。更多还原

【Abstract】 Objective Image dehazing is a critical and challenging task in computer vision, aimed at restoring clear images from hazy inputs affected by atmospheric scattering. While deep learning-based methods, particularly convolutional neural networks(CNNs), have shown promising results, existing approaches often struggle to balance the capture of both global contextual information and finegrained structural details necessary for effective dehazing. Moreover, many state-of-the-art models suffer from high computational complexity and large parameter counts, limiting their practical deployment. This paper proposes a novel gated fusion network integrating multi-scale feature extraction and cooperative attention mechanisms to achieve high-quality dehazing with maintained efficiency.Methods We propose a gated fusion network with multi-scale and cooperative attention(GFN-MSCA), an end-to-end trainable encoder-decoder architecture. The network incorporates two key innovative modules: the spatial-channel cooperative attention(SCSA) module and the multi-scale feature fusion(MSFF) module. The SCSA module employs a gating mechanism to adaptively recalibrate features by modeling interdependencies between spatial and channel dimensions, allowing the model to focus on information-rich regions and features. The MSFF module utilizes parallel convolutional branches with different receptive fields(1×1, 3×3, 5×5, 7×7) to extract features at various scales, which are then fused and refined using the SCSA module to handle diverse haze patterns effectively. These modules are embedded within a deep network composed of multiple residual groups and a cross-group feature fusion strategy, enhancing feature aggregation and gradient flow. The overall network architecture is illustrated. The model uses L1 loss for pixel-level accuracy to ensure high visual quality. Training and evaluation are conducted on the RESIDE benchmark dataset, including both indoor(ITS) and outdoor(OTS) subsets, with performance measured using peak signal-to-noise ratio(PSNR) and structural similarity index(SSIM).Results and Discussions Extensive experiments on the RESIDE SOTS datasets demonstrate the superior performance of the proposed GFN-MSCA method. On the indoor test set, our model achieves a PSNR of 38.86 dB and an SSIM of 0.9912, outperforming many compared state-of-the-art methods while maintaining a relatively low parameter count of only 4.99×10~6. On the outdoor test set, it attains a PSNR of 34.63 dB and an SSIM of 0.985, indicating strong generalization capability. Qualitatively, the proposed method produces visually compelling results with enhanced color saturation, preserved fine details, and minimal artifacts. GFN-MSCA effectively removes haze without introducing color distortion or oversaturation, closely matching the ground truth images in both indoor and challenging outdoor scenarios. Ablation studies further validate the contribution of each component. Replacing the basic spatial attention with the proposed SCSA module, coupled with the MSFF module, leads to a significant performance gain, confirming the importance of cooperative spatial-channel modeling and multi-scale feature fusion. The results indicate that the gated fusion mechanism successfully adapts to varying haze densities and distributions, enabling robust and naturallooking dehazing across diverse scenes.Conclusions This paper presents a novel GFN-MSCA network for single image dehazing that effectively integrates multi-scale feature extraction with a cooperative attention mechanism. The introduced SCSA module adaptively fuses channel and spatial information via a gating strategy, while the MSFF module captures and aggregates features at multiple scales to handle global haze context and local details simultaneously. Extensive experimental results on benchmark datasets confirm that the proposed method achieves great performance in both quantitative metrics and visual quality, while maintaining a compact model size. The network demonstrates excellent robustness and generalization across various haze conditions, making it a suitable solution for practical image restoration applications.更多还原